Skip to content

TestCorpus.md

Latest commit

 

History

History
3353 lines (2828 loc) · 193 KB

TestCorpus.md

File metadata and controls

3353 lines (2828 loc) · 193 KB
Raw

Local Test Corpus

Status: local developer machine only; not part of the repository. Last inventoried: 2026-10-30

This document records the legally obtained BIOS and disc images available on the development machine for manual compatibility, boot, audio, input, save-state, and performance testing. None of these files may be committed, packaged, or uploaded.

  • BIOS directory: bios/ in this repository (ignored by .gitignore)
  • Disc directories: machine-specific, set FIVEDO_GAMES to the main root:
    • Main collection: J:\Panasonic 3DO (589 CHDs)
    • MAME software-list collection: G:\MAME\MAME Software List CHDs\3do (5 CHDs)
    • Older paths used in earlier handoffs: D:\Samples\3DO and E:\Panasonic 3DO

All locations are outside version control. bios/ is already ignored by .gitignore, and *.chd/*.iso/*.bin/*.cue are ignored as a second guard. The full-collection sweep uses both roots explicitly; see CompatibilitySweep.md and GameCompatibilityList.md. The corpus root is never hard-coded in scripts; use $env:FIVEDO_GAMES and the -Roots parameter so the same commands work on other machines.

The corpus contents differ between machines. Do not treat a title that is missing on one machine as unavailable in the project. Use tools/release-soak.ps1 to discover the local discs, run the 36,000-frame release soak for the titles that are present, and record absent titles as Missing on this machine. If an archived 12k state is available for an absent disc, the script can continue it to the target frame count; that is continuation evidence, not a fresh-boot check.

2026-10-27 audio-clock rebaseline. The DSP sample clock now produces exactly 44,100 samples per emulated ARM second (see PortingNotes.md 12.4). Hashes recorded before this date were produced by FreeDO's 8.24 sample accumulator, which ran ~0.7% fast. Audio content and spectral peaks are unchanged, but FMV/loading frames can be one or more frames apart. The current 3,000-frame keys are in section 1.1; longer-run hashes in this document are historical unless they are explicitly marked as post-audio-clock.

1. Disc images

Title File Format Size (bytes) Game ID Known fixes
Alone in the Dark (USA) Alone in the Dark (USA).chd CHD 184,385,537 9f7d72bc FIX_BIT_TIMING_6
Lemmings (USA) Lemmings (USA).chd CHD 299,910,255 0be38b22 none in the 4DO database
Road Rash (USA) Road Rash (USA).chd CHD 506,330,721 d7672550 none in the 4DO database
Space Ace (USA) Space Ace (USA).chd CHD 229,432,527 112b53d0 none in the 4DO database
Wolfenstein 3D (USA) Wolfenstein 3D (USA).chd CHD 58,013,447 3dfcdbbf none in the 4DO database

All five are CHD images. They are the primary real-content corpus for:

  • BIOS + disc boot validation.
  • CHD sector streaming and read-ahead.
  • Title/menu rendering and input.
  • DSP/audio validation.
  • In-game save-state roundtrips.
  • End-to-end performance and frame-pacing measurements.

Seven additional CHD titles (Gex, Need for Speed, Super Street Fighter II Turbo, Total Eclipse, Killing Time, Theme Park, and Return Fire) were validated for boot, render, save-state, and audio in section 1.7.

The full local collection is inventoried in section 1.1bh: 1,371 images collapse to 414 unique titles and a 378-title game test set, with canonical disc selection and a 5DO/4DO/Phoenix comparison. Use artifacts/corpus/unique-testset.txt for compatibility runs; the five-title list above remains the differential corpus.

ISO/BIN/CUE parity testing still needs a legally obtained image in one of those formats, or a locally generated fixture. Do not convert and commit a retail disc.

1.1 Current boot baseline

The following results were re-recorded on 2026-10-30 after the headless PBUS fix, with panafz1.bin, a Release build, and run --frames 3000 --audio null. The CLI now attaches a default player-1 pad, matching the desktop shell; the earlier keys were recorded while headless runs reported no connected controller and are kept below for historical comparison. The CLI loads a fresh 32 KB NVRAM volume by default; pass --nvram <path> to persist it. They are the current regression baseline, not a full compatibility pass:

Title Final PC Frame hash Frame at 3,000 Observation
Alone in the Dark 0x000C4C94 0x05225F76 boot/logo AudioFolio init succeeds; with input and NVRAM it reaches the in-game room/inventory scene by 10,000 frames
Lemmings 0x0012B324 0x9C2E3236 rendered Reaches CyberIntro; at 10,000 frames with input it reaches the Load/Options title menu
Road Rash 0x000BA0F4 0x5E9EE515 rendered Loads the game OS (grafmath 21.10.504); SAudio init succeeds; the 3,000-frame capture is an intro-FMV frame; the 30,000-frame run reaches the 3D race scene
Space Ace 0x00076418 0xD666228E rendered ReadySoft player finds the CD-ROM, plays the demo, and passes the former frame-681 crash; the 3,000-frame capture is the adjacent demo frame
Wolfenstein 3D 0x00031F74 0x9A3B5131 boot/logo Loads Sherry/Operator/FS v24.225; differs from the pre-clock frame only in the star-field sparkle; with input it reaches gameplay (FLOOR 1-1)

The pre-PBUS-fix 2026-10-27 keys were (0x4EC2E1A7, 0x6CCA83F5, 0xCE90FA5A, 0xF349263B, 0x9A3B5131). The 2026-10-24 keys were (0xF4BDF79B, 0x932C7722, 0xEB31636F, 0xA2B49CF0, 0xC21BB6A1); the immediate pre-clock build at commit 46785b2 already produced 0x4EC2E1A7/0x6CCA83F5 for Alone/Lemmings because of the VDLP pixel-stride fix. Do not use the older keys after this date. The PBUS fix changes input-sensitive hashes because games now see a connected pad; re-record longer-run hashes before using them as regression keys.

1.1a ARM timing differential (2026-10-31; class ledger 2026-11-04)

The CORE-01 option was checked on the Lemmings boot with the same BIOS, disc, frame count, and fresh NVRAM. The FreeDO-compatible default is unchanged; the ARM60 datasheet mode runs more instructions per frame because its N cycle is two ARM clocks instead of four, so the guest reaches a different point at 3,000 frames.

Mode Final PC Frame hash Instructions Cycles/frame
--arm-timing freedo 0x0012B324 0x9C2E3236 194,135,650 208,356
--arm-timing arm60 0x0012A324 0xFBED38D3 272,410,816 208,341

Both modes completed the boot and produced a rendered frame. This is a single-title smoke differential, not the full-collection report required to change the default.

The same pair of runs was repeated with --cycle-classes to audit the per-class ledger introduced for CORE-01. The final frame hashes match the 2026-10-31 table above (0x9C2E3236 and 0xFBED38D3), and the class averages show the expected N=2 reductions:

Class freedo cycles/instruction arm60 cycles/instruction
Multiply 8.21 7.87
Swap 10.00 6.00
DataProcessing 1.06 1.03
SingleDataTransfer 6.38 3.98
BlockTransfer 10.07 7.42
Branch 4.18 3.11
SoftwareInterrupt 6.00 4.00

Undefined and Coprocessor did not occur during this boot. Every executed class is within the exact per-instruction table in PortingNotes.md 12.28; the averages move only because of operand-dependent multiply cost and condition-failed one-cycle slots. The differential smoke (benchmarks/diff/smoke.json, 300 frames, accuracy profile) was re-run after the ledger change and still reports Match with frame hash 0x110E7DC5 on both engines.

1.1a1 40-title ARM timing A/B (2026-11-04)

The CORE-01 option was measured across the 40-title sample (benchmarks/diff/sample40.json, 300 frames, one checkpoint, --no-game-fixes, cdSpeed: unbounded). Both 5DO runs used the same staged ISO that the Opera reference captured from, the same BIOS, and a fresh NVRAM. The accuracy profile supplied the CLIO/CEL/CD timing; only the ARM mode changed, selected by the new harness override --arm-timing freedo|arm60.

ARM mode Matched Diverged Errors
freedo (N=4) 15 25 0
arm60 (N=2) 17 23 0

Only two of the 40 final frame hashes changed, and both moved from Diverged to Match:

Title freedo hash arm60 hash Opera hash
20th Century Video Almanac (USA) 0x6CDFBDC5 0x110E7DC5 0x110E7DC5
Battle Chess (USA) 0x6CDFBDC5 0x110E7DC5 0x110E7DC5

The other 38 hashes were identical. This is a 300-frame boot sample, not the 3,000-frame 594-disc report, so the 2026-11-04 decision keeps ARM60 opt-in. It does show that the N=2 mode is not a regression risk on this sample and improves parity where the ARM instruction stream reaches the frame boundary first. Command:

tools/FiveDo.DiffHarness/bin/Release/net10.0/FiveDo.DiffHarness.exe `
  run artifacts/diff/sample40-iso.json --out artifacts/diff/sample40-arm60 `
  --arm-timing arm60 --no-images --resume

sample40-iso.json is the local sample40 manifest with each disc/iso pointing at the staged ISO under artifacts/diff/work/discs; it is not committed because the paths are machine-specific. The staged ISO is the data-sector image used by the reference core, so CD-DA and subchannel behavior are outside this boot sample. The Opera captures are reused by --resume, so the second run differs only in the 5DO ARM mode.

1.1b CLIO timer scheduler differential (2026-10-31)

The CORE-02 option was checked on the 3,000-frame boot with the same panafz1.bin, fresh NVRAM, and disc files. The FreeDO-compatible default is unchanged; the Opera per-slot scheduler changes when timer FIQs fire, so the guest reaches a different instruction boundary at 3,000 frames.

Title Mode Final PC Frame hash Instructions Cycles/frame
Lemmings --clio-timers freedo 0x0012B324 0x9C2E3236 194,135,650 208,356
Lemmings --clio-timers opera 0x0012DA44 0x44EEA93D 192,399,552 208,343
Gex --clio-timers freedo 0x001105C4 0x05F38B50 209,334,971 208,338
Gex --clio-timers opera 0x00110524 0xA9DE7B03 209,116,136 208,335
Road Rash --clio-timers freedo 0x000BA0F4 0x5E9EE515 171,049,511 208,342
Road Rash --clio-timers opera 0x00098AF0 0x36359402 170,826,129 208,350

All six runs completed and rendered. This is a smoke differential, not the full 594-disc report required to change the default. The Lemmings and Road Rash FreeDO hashes match the section 1.1 keys, confirming that the new option did not alter the default path.

1.1b1 40-title CLIO timer A/B (2026-11-04)

The 40-title sample (benchmarks/diff/sample40.json, 300 frames, one checkpoint, accuracy timing, --no-game-fixes, cdSpeed: unbounded) was captured twice with the same build and the same Opera reference captures. Only the CLIO mode changed: the accuracy timing selects OperaScheduler; the second run used the new harness override --clio-timers freedo, which is forwarded to every 5DO capture.

CLIO mode Matched Diverged Errors
OperaScheduler 17 23 0
FreeDoCompatible 17 23 0

The same 17 titles match in both modes, and all 40 final frame hashes are identical. The mode changes the checkpoint PC for 37 of the 40 titles but not the rendered frame at 300 frames, so the CLIO scheduler alone neither closes nor regresses a title in this boot sample. The accuracy baseline is 17/40 rather than the previously recorded 18/40 because the CEL pixel-accounting fix changed the 5DO accuracy hashes; the Opera reference captures are unchanged. The mode stays opt-in pending the full 594-disc 3,000-frame run.

Timer ledger and trace evidence: a 300-frame Lemmings boot with --clio-timers opera --clio-timer-stats --clio-timer-trace reports timer 0 (141,814 ticks, 2 reloads), timer 3 (70,688 ticks, 55 stops and 55 FIQs), and timer 4 (37,600 ticks, 93 reloads). The trace shows timer 14 reloading 0x3d every ~12,400 ARM cycles at slack 0x150: one scan is 15 two-cycle slots plus the (4 + 336) M25M-tick gap, or 200 ARM cycles, so 62 executions take 12,400 cycles. The same title with --clio-timers freedo reports 183,177 timer-0 and 95,104 timer-3 ticks. That is the expected FreeDO all-timer scan rate: 25 MHz/64 = 390.6 kHz per timer, versus Opera's 195.3 kHz at the default slack of 64 (62.5 kHz once the guest writes the 0x150 slack used by this title).

1.1c CEL busy timing differential (2026-10-31)

The CORE-03 option was checked on the 3,000-frame boot with the same panafz1.bin, fresh NVRAM, and disc files. The FreeDO-compatible default is unchanged. Opera's busy model defers the CEL draw and keeps the status register busy, so titles that render large CELs early reach a different instruction boundary at 3,000 frames.

Title Mode Final PC Frame hash Instructions Cycles/frame
Lemmings --cel-timing freedo 0x0012B324 0x9C2E3236 194,135,650 208,356
Lemmings --cel-timing opera 0x0012B324 0x9C2E3236 194,135,650 208,356
Gex --cel-timing freedo 0x001105C4 0x05F38B50 209,334,971 208,338
Gex --cel-timing opera 0x001105C4 0x05F38B50 209,334,971 208,338
Road Rash --cel-timing freedo 0x000BA0F4 0x5E9EE515 171,049,511 208,342
Road Rash --cel-timing opera 0x000B9E9C 0xA39F031F 170,868,164 208,338

All six runs completed and rendered. Lemmings and Gex are pixel-identical because their early boot path does not poll CEL status; Road Rash diverges at its early CEL-heavy boot frames, which is the expected effect of serializing CEL work. This is a smoke differential, not the full 594-disc report required to change the default.

Four more CEL-heavy titles completed 3,000 frames with --cel-timing opera:

Title Final PC Frame hash
Alone in the Dark 0x000C4CB4 0x328ABCD3
Space Ace 0x00076458 0xD666228E
Wolfenstein 3D 0x000320D8 0x9A3B5131
Total Eclipse 0x000C9870 0xAD0C92D2

Lemmings and Road Rash also completed 12,000 frames with the Opera model (0xCE7E7A74 and 0xAEE9963B respectively), so the deferred draw and busy deadline do not stall those boot paths.

1.1c1 CEL busy work accounting (2026-11-04)

The 2026-10-31 runs used a work ledger that missed the 320x240 literal-line and packed-repeat rasterizer paths: those paths wrote through CelMemory.WriteHalfword without calling CelEngine.AddPixels, while the high-resolution paths went through WritePixel and were counted. Opera counts every process_pixel call and its direct TexelDraw_Line writes, so the busy deadline was too short for those lists.

After the fix, a 3,000-frame Road Rash boot with --cel-timing opera --cel-trace reports five full-screen lists as cels=1 pixels=76800 busy=106730 (two of them) and cels=2 pixels=153600 busy=213461; the literal-line lists previously reported zero pixels. The FreeDO path still produces the remembered golden frame hash 0x5E9EE515 (PC 0x000BA0F0 on this fresh-NVRAM run; instruction totals vary with the initial NVRAM bytes, so the frame hash is the regression key). The Opera path is now:

Title Mode Final PC Frame hash Instructions
Road Rash --cel-timing opera 0x000B9D54 0x9512219E 170,818,687
Lemmings --cel-timing opera 0x0012B324 0x9C2E3236 194,135,071
Gex --cel-timing opera 0x001105C4 0x05F38B50 209,334,014

Road Rash also completed 12,000 frames in both modes (PC 0x00020640 for each): FreeDO retained 0x2D05A53C and Opera produced 0x81BDF30F, so the longer busy periods do not stall the boot. Lemmings and Gex do not poll CEL status in their early boot, so they remain pixel-identical to the FreeDO path. The 300-frame accuracy-profile smoke still matches the reference (0x110E7DC5 on both engines). The work-accounting change is covered by PackedLine_CountsDrawnPixelsForOperaBusy, PackedRepeatLine_CountsEveryRetainedPixelForOperaBusy, and LiteralLine_CountsDrawnPixelsForOperaBusy.

A six-title, 3,000-frame, accuracy-profile differential on 2026-11-04 completed both engines with no capture errors; all six still diverged at the final checkpoint. The accounting fix changes 5DO's CEL timing but does not by itself close the remaining scheduling/render gap, so the Opera busy mode stays opt-in.

Title 5DO frame hash Opera frame hash
Road Rash 0xACFEC90C 0x3F570FC0
Gex 0xD92DBD3C 0xBED174BD
AutoBahn Tokio 0x04C4738C 0x32A050A8
Johnny Bazookatone 0x5796BF92 0xA29AD2D1
Primal Rage 0xCD9812F9 0x8E116314
DinoPark Tycoon 0x266F986F 0xB45E532B

1.1d CD-ROM cycle timing differential (2026-10-31)

The CORE-04 option was checked with panafz1.bin, fresh NVRAM, and the same discs. The FreeDO-compatible default is unchanged. The Opera cycle model schedules each 2048-byte data sector at the stock 2x drive speed (81,246 ARM cycles at 12.5 MHz after Opera's 167 us overhead compensation), holds the data poll clear until the sector is ready, and publishes the read-data status after the transfer is drained. The XBUS CD FIQ is raised when the delayed data or status becomes ready.

Title Frames Mode Final PC Frame hash Observation
Lemmings 3,000 --cd-timing freedo 0x0012B324 0x9C2E3236 unchanged baseline
Lemmings 3,000 --cd-timing opera 0x00059710 0xB1CE635A still loading; expected slower CD
Lemmings 12,000 --cd-timing freedo 0x00128490 0xCE7E7A74 title/menu
Lemmings 12,000 --cd-timing opera 0x0012918C 0x110E7DC5 boot reaches the intro later
Lemmings 24,000 --cd-timing opera 0x0012DE58 0x9D4DBD37 intro animation rendered
Road Rash 12,000 --cd-timing freedo 0x00020640 0x2D05A53C intro credits
Road Rash 12,000 --cd-timing opera 0x00020640 0x37D5CCFE intro credits at a different FMV point
Gex 12,000 --cd-timing freedo 0x0008D3BC 0x9B1F45AE boot/menu
Gex 12,000 --cd-timing opera 0x000820F8 0x259870E0 boot/menu, later CD phase

All nine runs completed and rendered. The mode is intentionally slower than the immediate FreeDO path because it models the real 2x drive; the important result is that the boot, OS, and FMV paths continue rather than stalling. This is a three-title smoke differential, not the full 594-disc report required to change the default. The five-title 3,000-frame default regression keys are unchanged from section 1.1 (Alone 0x05225F76, Lemmings 0x9C2E3236, Road Rash 0x5E9EE515, Space Ace 0xD666228E, Wolfenstein 0x9A3B5131), confirming that the new option did not alter the FreeDO path.

At 10,000 frames with input and a fresh NVRAM (pre-transposition-fix):

Title Final PC Frame hash Observation
Alone in the Dark 0x0004B6B4 0x67FD3C6F In-game room/inventory scene
Wolfenstein 3D 0x0001F014 0xA5A33C4D Gameplay (FLOOR 1-1)

These two hashes are retained for historical comparison; the current first-pass numbers are in section 1.2.

1.1e CD speed and timing-profile differential (2026-10-31)

The SYS-09 speed option was checked with panafz1.bin, fresh NVRAM, and 3,000 frames. The FreeDO path is independent of --cd-speed: Lemmings with freedo/X2, freedo/X1, and freedo/Unbounded all end at PC 0x0012B324, hash 0x9C2E3236, matching section 1.1. The cycle-based path changes with the speed setting, and every run completed and rendered:

Title Speed Final PC Frame hash CD poll
Lemmings 1x 0x0012B1C0 0x50A882FC 0x4B
Lemmings 2x 0x00059710 0xB1CE635A 0x49
Lemmings 4x 0x00012A08 0xB5A493B9 0x4B
Lemmings 8x 0x0012DE7C 0x6D9321A4 0x49
Lemmings 32x 0x0012A308 0x1FCCAB07 0x49
Lemmings unbounded 0x0012B42C 0x1F4E236C 0x49
Gex 1x 0x00013F7C 0xB87FA1D5 0x0F
Gex 2x 0x0010FD60 0x496B0DFE 0x0F
Gex 4x 0x0002A7D4 0x2BF57DC4 0x0D
Gex 8x 0x0000E19C 0xCF03CA3B 0x0C
Gex unbounded 0x0010F908 0x05F38B50 0x0C
Road Rash 1x 0x000B9D08 0x8733E5EB 0x0F
Road Rash 2x 0x0001276C 0x6F93310D 0x0C
Road Rash 4x 0x000B9D80 0xB2B1325E 0x0F
Road Rash 8x 0x000BA0E0 0xA64C80AC 0x48
Road Rash unbounded 0x000B9C40 0x7EBB3876 0x48

The CD poll column is diagnostic only; the 0x40 media-access pulse is expected on the idle READ_ID paths. This is a three-title smoke differential, not the full-collection report required to change the default timing model. The --profile accuracy path was also smoke-tested through the CLI: it reports profile: Accuracy, armTiming: Arm60Datasheet, clioTimers: OperaScheduler, celTiming: OperaBusy, cdromTiming: OperaCycleBased, and cdSpeed: X2; the 3,000-frame Lemmings run reaches PC 0x000109AC, hash 0xF303CBB4. An explicit --arm-timing freedo override reports profile: Custom and the FreeDO timing value. The individual option tests and state-mismatch tests are in tests/FiveDo.Tests.

1.1f BIOS manager and auto-selection (2026-10-31)

bios scan bios finds 8 primary BIOS images, 2 secondary images (the FZ-1 Kanji font and the 512 KB arcade support ROM), and 0 errors. The common Panasonic FZ-1/FZ-10/ FZ-10E/FZ-10JA, GoldStar, Sanyo, patched arcade, Kanji font, and arcade support hashes are recognized by MD5; unknown images are classified from the file name and marked unverified.

The five-title 3,000-frame boot corpus was rerun with --bios-auto bios, a fresh NVRAM, and the Release build. Automatic selection chose the FZ-10 for these US titles; the hashes differ from the FZ-1 section 1.1 keys because the BIOS image differs, but every run completed and the final frame was non-black.

Title Selected BIOS Final PC Frame hash
Alone in the Dark (USA) Panasonic FZ-10 0x00021BEC 0x175B8FE4
Lemmings (USA) Panasonic FZ-10 0x0012DA2C 0x0C5B2F8E
Road Rash (USA) Panasonic FZ-10 0x000BA13C 0x6D23530B
Space Ace (USA) Panasonic FZ-10 0x0007A160 0xF55471B4
Wolfenstein 3D (USA) Panasonic FZ-10 0x00018FAC 0x175B8FE4

Known-title requirements were checked at 12,000 frames. These are the BIOS choices the BIOS-aware triage in Compatibility.md 13.2 identified; each final frame contained rendered content.

Title Selected BIOS Final PC Frame hash
Powers Kingdom (Europe) Panasonic FZ-10E (Anvil) 0x00010F3C 0xD5F7AF94
Slayer (USA) Panasonic FZ-10 0x000EC8B8 0xA50A108B
Slayer (Europe) Panasonic FZ-1 0x00020640 0xAB8BC516
Crime Patrol (USA) (Arcade CD-ROM) Panasonic FZ-10 (patched) 0x00070D20 0x76C2D1D1

This is smoke evidence, not the full-collection BIOS-aware sweep. The configured BiosPath remains the fallback and is still used when BiosFolder is empty, so existing settings keep their exact BIOS.

1.1g Per-game NVRAM smoke (2026-10-31)

Two titles were run at 3,000 frames with panafz1.bin, a fresh temporary NVRAM root, --nvram-mode per-game --nvram-version 1, and --audio null. Each run created its own 32,768-byte image and left the other title untouched:

Title Final PC Frame hash NVRAM file
3D Atlas (USA) 0x000189C4 0x1187598D 83d70d9c6a4a525a6774ae22ddb1fa23.1.ram
Lemmings (USA) 0x0012B324 0x9C2E3236 0be38b22987b40b5d48f70a167656049.1.ram

Lemmings' hash matches section 1.1, confirming that per-game storage does not change emulation. nvram list --dir <root> --json reported both images with the correct game IDs, modes, versions, and lengths. The store and CLI tests cover shared-to-per-game migration, version separation, corrupt-primary backup, and import/export/reset; see tests/FiveDo.Tests/Host/NvramLayoutTests.cs, NvramStoreTests.cs, and tests/FiveDo.Tests/Cli/CliProcessTests.cs.

The 30,000-frame Road Rash run after the CCB fix originally ended at PC 0x00016194 with frame hash 0xE2D08131 and rendered the race scene when run without NVRAM. That result was produced before the arbitrary-CEL transposition fix; the transposed CELs were corrupting the OS VRAM free list, so the "frozen race" was actually a failed allocator. After the fix the same run reaches 60,000 frames with PC 0x006940C0-class game code, DSP running, no AllocMem error, and a correct race render. See section 1.2 for the new baselines.

Use the frame hash as the stable regression key and record the PC alongside the commit. The pre-NVRAM 3,000-frame hashes were (0x85EF87FF, 0x932C7722, 0xEB31636F, 0x2A9967FA, 0x1A940707); the pre-CCB hashes were (0x85EF87FF, 0x932C7722, 0xEB31636F, 0x6AA8B441, 0x1A940707). The NVRAM fix changed Wolfenstein, Alone, and Space Ace because their save-space checks now see a valid volume; the CCB fix changed Space Ace's rendering. The 3,000-frame hashes are unchanged by the transposition and VDLP fixes because those frames are all boot/FMV screens.

PC 0x205C4/0x205C8 is the Portfolio idle loop, not a halt: VDL and timer interrupts continue to run and the boot log advances. Earlier baselines that described this address as a "halt path" were misreading the idle loop.

1.1g Differential harness (2026-10-31)

The first automated 5DO-vs-OperaCore run uses tools/FiveDo.DiffHarness and benchmarks/diff/smoke.json:

  • Title: Lemmings (USA), panafz1.bin, 300 frames.
  • 5DO: --profile accuracy --cd-speed unbounded --no-game-fixes, checkpoint 300, frame hash 0x110E7DC5.
  • OperaCore: same disc data (cached ISO), same BIOS, 300 frames, checkpoint 300, frame hash 0x110E7DC5.
  • Result: Match. 5DO ran 300 frames in 1.3 s; OperaCore ran in 5.1 s.
  • The two engines also produced the same final frame hash at 300 frames but diverged at frames 100 and 200: Opera reached the Lemmings boot screen earlier. That is a real scheduling difference, not a renderer difference. The smoke manifest now checkpoints only the matched frame; the early-frame divergence is retained in the run history and will be triaged with CORE-01/02/04 evidence.

The five-title benchmarks/diff/showcase.json run (3000 frames per title) is the next evidence set. It compares frame hashes only; the reference core seeds its PRNG from wall-clock time, so audio hashes are recorded but not part of the pass/fail gate. See DifferentialHarness.md for the manifest, staging rules, allowlist policy, and limitations.

First showcase result (2026-10-31, 3,000 frames, checkpoint 3,000):

Title 5DO OperaCore Result
Lemmings (USA) 0x20C4E404 0xEBDB01CA Diverged
Gex (USA, Europe) 0xC8A2B60F 0xBED174BD Diverged
Road Rash (USA) 0x6D066B83 0x0528251F Diverged
Wolfenstein 3D (USA) 0x175B8FE4 0x175B8FE4 Match
Alone in the Dark (USA) 0x175B8FE4 0x175B8FE4 Match

The two matches are exact pixel matches at the same frame. The three divergences are at game-code PCs (0x0012DE84, 0x0011005C, 0x000B9FAC) with the same CD status class, so they are candidates for the CORE-01/02/04 timing differences, not a failure to load. The next pass adds frameWindow: 2 diagnostics and checkpoint images so the report can distinguish "same pixels one frame apart" from a real pixel/state divergence.

Strict PC gate (debug core, same captures):

Title Video PC delta (Opera - 5DO)
Lemmings (USA) Diverged -11,076
Gex (USA, Europe) Diverged 176
Road Rash (USA) Diverged 284
Wolfenstein 3D (USA) Match -184,696
Alone in the Dark (USA) Match 739,184

With --compare-pc, every title is a divergence: Wolfenstein and Alone render identical pixels at frame 3000 while executing completely different code paths. This is the central lesson for CORE-16: a frame-hash match is necessary but not sufficient. Use the video gate for renderer/CLUT regressions and the PC gate (with the debug core) for instruction-stream timing.

Checkpoint images (second pass, 2026-10-31): the harness now writes f003000.png for 5DO and f003000.bmp for Opera. The images confirm that the divergences are state/animation differences, not palette or scaling bugs:

  • Lemmings: the same intro bedroom scene, but the character's pose differs (6,160 of 76,800 sampled pixels differ); the difference is stable across Opera frames 2998-3000, so it is not a simple ±1/2-frame offset.
  • Gex: a different camera/character pose in the same opening scene.
  • The ±2-frame frameWindow did not find an exact hash match for either title, so the harness correctly reports Diverged, not Offset.

The next triage step is to compare the 5DO CEL/CD/CLIO traces at frame 3000 with the Opera reference and determine whether the divergence starts in the CD load handshake, CEL busy timing, or the ARM/timer scheduler.

1.1h SYS-15 input devices (2026-11-01)

Lemmings (USA), panafz1.bin, 300 frames:

Configuration Frame hash PBUS packet (first frame)
Legacy all-joypad 0x110E7DC5 (286,162,373) 00480080000000000000000000800000
--pbus-format opera 0x110E7DC5 (286,162,373) 8000FFFFFFFF...
--input-device 1=mouse 0x110E7DC5 (286,162,373) 49000000FFFFFFFF...

The chain and legacy layouts render the same boot frame. Exact encoder tests cover the joypad, flightstick (01 7B 08 signature), mouse (0x49), lightgun (0x4D, 20-bit scan counter and 0x1F hit count), arcade lightgun, Orbatak trackball (0x49), and Orbatak buttons (0xC0), including 0xFF padding and mixed-device port order.

Representative arcade titles, 3,000 frames, panafz10-patched.bin:

Title Device Frame hash PC Image
Crime Patrol (USA, Arcade CD-ROM) arcade-lightgun 0xFEC54535 (4,274,341,173) 0x0008F9E0 Non-black, differs from joypad
Crime Patrol (USA, Arcade CD-ROM) joypad 0x0430CFAD (70,307,757) 0x0008F744 Reference
Mad Dog II (USA, Arcade CD-ROM) arcade-lightgun 0xC5F4A77C (3,321,145,212) 0x00090554 Non-black
Orbatak (USA, Arcade CD-ROM) orbatak-trackball 0x9BCA1E9E (2,613,714,590) 0x000A7428 Non-black, differs from joypad
Orbatak (USA, Arcade CD-ROM) joypad 0x322D23AB (841,819,051) 0x00083570 Reference
Shootout at Old Tucson (USA, RE3) arcade-lightgun 0x9F199B5A (2,669,255,514) 0x00020600 Non-black

The different hashes and PCs between the joypad and peripheral runs show that the games detect the selected device. These are boot/title checks, not scripted playthroughs; calibration and gameplay input remain.

Optional debug-core fields (2026-10-31): a locally built OperaCore with the read-only sg_debug_* exports records the reference PC, ARM state hash, and CD FIFO status. On the Lemmings 300-frame smoke the frame hash matches while the PC delta is 4 bytes (0x000205F8 vs 0x000205FC). At frame 3000 the delta is 0x2B44 (11,076 bytes): 5DO 0x0012DE84, Opera 0x0012B340. That is a real instruction-stream divergence, not a renderer bug. The CD FIFO status/poll values are reference-internal and are captured for comparison but are not the same fields as 5DO's DriveStatus/Poll; do not compare them directly.

Opera parity batch 1 (2026-10-31): the first source-level parity batch (VDLP next/force-line, ARM immediate rotate carry, DSP CPU FIFO, MADAM span/MARIA/BPP0/SKIPX/blue-LSB/REGCTL/PIXC, CLIO FIFO) was validated with the full unit suite and the differential harness:

  • dotnet test 5DO.sln: 1,140 passed, 0 failed.
  • Smoke: Lemmings (USA), 300 frames, checkpoint 300: PASS, 1 match, 0 diverged, 5DO 0x110E7DC5 = Opera 0x110E7DC5.
  • Showcase: 3,000 frames, checkpoint 3,000: unchanged from the pre-batch baseline; Wolfenstein and Alone still match, Lemmings/Gex/Road Rash still diverge, and the 5DO hashes are identical to the pre-batch values.
  • VDLP save states are now version 3; v1 and v2 states migrate to the head/successor cursor model.

OP-14 retirement (2026-10-31): the automatic GraphicsStepY (Samurai Shodown) and BitTiming6 (Alone in the Dark) mappings were removed. The enum values remain loadable for old states, but they are no-ops. The literal-CEL path now always reads the actual source row.

  • Samurai Shodown (USA/EU): frames 3,000/4,000/5,000/6,000 are identical with and without GraphicsStepY.
  • Alone in the Dark (USA): with BitTiming6 disabled, the credits use the actual row and no longer stretch vertically. A 9,000-frame harness probe still diverges from Opera for unrelated CD/timing reasons.
  • DinoPark Tycoon (USA): the 3,000-frame image renders correctly with the generic XYPOSH/XYPOSL origin model and its existing BitTiming3 scale fix.
  • VDLP now keeps the retained head separate from the active successor cursor and processes a zero-persist successor on the same line. The Opera scanline-5 boundary is deferred until OP-SYS-05 aligns 5DO's visible start line (16) with Opera's (21); the line-0 cadence keeps Lemmings at 0x110E7DC5.

OP-SYS-02 CD poll/local reset (2026-10-31): the accuracy timing mode now uses Opera's poll semantics. POLREMASK (0x08) performs a local reset that clears the command/transfer FIFOs and sector position; POLRE (0x80) is a write-one-to-clear media-access latch; status/data/command access is gated during reset; eject reports READY without the FreeDO phantom disc bits; and a loaded state with reset asserted runs the same reset. CD-DA read-ahead and the READ_ID heuristic remain for SYS-01.

  • dotnet test 5DO.sln: 1,143 passed, 0 failed.
  • Differential smoke (Lemmings, 300 frames): PASS, 0x110E7DC5.
  • Five-title showcase: unchanged; all 5DO hashes identical to the pre-batch baseline.

OP-SYS-01 ROM overlay (2026-10-31): the memory bus now starts with the BIOS overlaid at address zero. Low reads return the active ROM bank; the first CPU write to DRAM clears the overlay; VRAM writes do not; reset reasserts it; and memory state version 2 saves/migrates the flag (v1 states load as post-boot). DIPIR latches, access-abort rules, and XBUS media access remain.

  • dotnet test 5DO.sln: 1,147 passed, 0 failed.
  • Differential smoke (Lemmings, 300 frames): PASS, 0x110E7DC5.
  • Five-title showcase: unchanged; all 5DO hashes identical to the pre-batch baseline.

OP-SYS-01 DIPIR/XBUS (2026-10-31): CLIO now has DIPIR1/DIPIR2 latches, the before-reset conversion, the DIPIR FIQ, and XBUS access-abort rules. The XBUS selected-device address is available to CLIO, Opera mode stops masking the poll high nibble, and poll/command writes can latch media access when the CD reports POLRE. The DIPIR/abort path is opt-in (EmulatorTuning.OperaDipir) because enabling it before OP-SYS-03's soft-reset/clear-DIPIR path changes the Lemmings boot; the default and accuracy profiles leave it off.

  • dotnet test 5DO.sln: 1,154 passed, 0 failed.
  • Differential smoke (Lemmings, 300 frames): PASS, 0x110E7DC5.
  • Five-title showcase: unchanged; all 5DO hashes identical to the pre-batch baseline.

OP-SYS-03 CSTAT soft reset (2026-10-31): CLIO now implements the CSTAT reset bits, clear-on-zero bits, DIPIR-reset bit, and a soft-reset callback. The machine handler performs an ARM soft reset (SPSR_svc/R14_svc preserved, PC=0, ROM overlay reasserted), MADAM/DSP soft resets, and optional XBUS controller/DIPIR clearing. The full path is opt-in (EmulatorTuning.OperaSoftReset) because enabling it before a boot transcript changes the Lemmings path; the register and ARM semantics are covered by unit tests.

  • dotnet test 5DO.sln: 1,157 passed, 0 failed.
  • Differential smoke (Lemmings, 300 frames): PASS, 0x110E7DC5.
  • Five-title showcase: unchanged; all 5DO hashes identical to the pre-batch baseline.

OP-SYS-06 PRNG split (2026-10-31): CLIO now has its own splitmix32 generator for 0x17F0 instead of borrowing the DSP LCG. The state is seeded from EmulatorTuning.RandomSeed when provided, saved in CLIO state version 3, and migrated from v1/v2. The DSP keeps its separate FreeDO-compatible LCG state.

  • dotnet test 5DO.sln: 1,159 passed, 0 failed.
  • Differential smoke (Lemmings, 300 frames): PASS, 0x110E7DC5.
  • Five-title showcase: 2 matches (Wolfenstein, Alone); Gex's hash changed because it reads 0x17F0, and its 3,000-frame image still shows a valid game scene. The other 5DO hashes are unchanged.

OP-SYS-10 CHD-zstd (2026-10-31): CHDSharp 1.6.0 already supports Zstd (ChdCodec.Zstd/Cdzstd) and ships VendoredZSTD.dll. Opening all 589 local CHDs succeeded; they use CDLZMA/CDZLIB/CDFLAC, so no zstd fixture is present. No native dependency or code change is required; add a zstd CHD fixture when one is available.

40-title differential sample (2026-10-31): benchmarks/diff/sample40.json runs the first 40 titles from the full manifest at 300 frames with the accuracy profile, no game fixes, unbounded CD speed, and no images. Result: 18 match, 22 diverged, 0 error; every 5DO run completed. This is the first broad post-parity baseline. The full 594-title collection remains runner-bound.

The next step is to capture a CLIO CSTAT/DIPIR boot transcript, compare the OperaDipir/OperaSoftReset path against Opera, and then enable it in the accuracy profile.

Opera-parity title probes (2026-10-31): 5DO completed 3,000 frames for Samurai Shodown (USA/EU), DinoPark Tycoon, Alone in the Dark, Johnny Bazookatone, AutoBahn Tokio, and Primal Rage with the accuracy profile. The stock OperaCore crashed on Samurai Shodown (0xC0000005) inside the harness; the other probes diverged at frame 3000, consistent with the existing PC/timing gap. A Johnny Bazookatone checkpoint image shows the game scene without the issue #291 horizontal-line artifact.

1.1h Rewind and fast-forward (2026-11-01)

Rewind and fast-forward are host features and do not change emulation output. The evidence is the unit/integration suite plus the state-equality test:

Test Result
RewindBufferTests.RewindTo_ReplaysInputAndRestoresTheExactMachineState PASS: a 90-frame run rewound to frame 37 produces a byte-identical save state to a fresh control run to frame 37.
RewindBufferTests.RewindTo_ClampsToTheOldestRetainedFrame PASS: targets before the oldest keyframe clamp to frame 0.
RewindBufferTests.Capture_ResetsHistoryWhenTheFrameCounterJumps PASS: a timeline jump clears stale keyframes.
RewindBufferTests.MemoryLimit_KeepsAtLeastOneKeyframe PASS: a one-byte cap retains one keyframe.
CoreRunnerTests.Rewind_RestoresAnEarlierFrameAndUpdatesTheRunnerCounter PASS: the runner and machine frame counters agree after rewind.
CoreRunnerTests.SpeedMultiplier_RunsEveryEmulatedFrameAndMutesAudio PASS: 4x runs every emulated frame, pushes no audio samples, and resets the audio queue.
FrameSkipStateTests.SkipFrame_OnlyChangesThePresentedFrameBuffer PASS: 60 frames with three of four rendered produce identical CPU, memory, MADAM, CLIO, DSP, SPORT, VDLP, XBUS, CD, Quartz, and options sections; only the Machine frame section differs.
GameDatabaseTests.AddMissingDefaultHotkeys_PreservesCustomBindings PASS: older profiles gain the Rewind/FastForward defaults without overwriting a custom save key.
HotkeyRouterTests.Resolve_ReturnsEveryDefaultHotkey PASS: R and Tab resolve to Rewind and FastForward.

Command:

dotnet test 5DO.sln -c Release --filter "FullyQualifiedName~RewindBufferTests|FullyQualifiedName~CoreRunnerTests|FullyQualifiedName~FrameSkipStateTests|FullyQualifiedName~HotkeyRouterTests|FullyQualifiedName~GameDatabaseTests"

The default timing profile and golden boot hashes are unchanged because the core machine is untouched; the rewind buffer only observes and restores complete states. The Lemmings (USA) 300-frame CLI smoke still reports 0x110E7DC5:

dotnet run -c Release --project src/FiveDo.Cli -- run `
  artifacts/diff/work/discs/lemmings-usa-smoke/lemmings-usa-smoke.iso `
  --bios bios/panafz1.bin --frames 300 --json

1.1i Disc browser and disc info (2026-11-01)

The disc browser is a read-only host model; it does not start the machine. Unit tests cover synthetic Opera images (tree order/depth, metadata totals, file/directory/all extraction, name sanitization, a real ISO file, and missing images). A manual probe against the local Lemmings CHD confirmed real CHD and database metadata:

Field Value
Source artifacts/chdmaker/lem.chd
Format CHD
Volume label CD-ROM
Game ID 0be38b22987b40b5d48f70a167656049
Sectors 206,126
Contents 461 files, 20 directories, 374,239,333 bytes
Database Lemmings, US, Psygnosis, 1993
Automatic fixes None
Extraction One selected file (181 bytes) written and removed

Command:

dotnet test tests/FiveDo.Tests/FiveDo.Tests.csproj -c Release --filter "FullyQualifiedName~DiscBrowserTests"

1.1j Multi-track, CD-DA, and volume (2026-11-01)

SYS-01/AUD-05 are covered by synthetic fixtures until a known CD-DA title is validated:

  • CueSheet_ExposesMultiTrackTableOfContentsAndAudioSectors builds one Mode 1 data track and one audio track, checks the logical start LBAs and native sector sizes, reads data through ReadSector, reads 2352-byte audio through ReadTrackSector, and verifies that an audio LBA has no 2048-byte user data.
  • ChdImage_ExposesAudioTracks encodes the same layout to CHD and checks the CHD track table plus byte-swapped audio samples.
  • DiscInfoReportsEveryTrackAndTheLeadOut, ModeSetToDigitalAudioReadsRawAudioSectors, PlayAudioTrackMovesTheSubQPositionAndStopsAtTheEnd, and ModeSetRejectsAnOversizedBlock cover the CD-ROM command path.
  • PlayAudioTrackMovesTheSubQPositionAndStopsAtTheEnd also verifies the first CD-DA sample is decoded little-endian from the audio sector.
  • DiscTracks_ReportsAnIsoImage also checks the --json track table.
  • CdAudioMixTests verifies that a full-volume CD-DA stream changes the packed DSP output, that mute produces the same output as zero volume, and that volume/mute roundtrip through the CD save-state section.
  • AppSettingsStoreTests covers FND-02 legacy migration, malformed-file backup, version roundtrip, invalid-value normalization, per-game profile roundtrip/apply, and Inspect/Migrate; the CLI tests cover settings validate|migrate.

The full suite is 1,381 tests in Debug. A Release 10-disc smoke after the track/TOC change (tools/compat-sweep.ps1 -Limit 10 -Frames 3000) reported 10/10 Loadable/Rendered, so the data-track path is unchanged. A real CD-DA title and a 10-second DSP+CD-DA WAV capture are still required for SYS-01 completion.

1.1k Multi-disc playlist and live swap (2026-11-01)

SYS-02 coverage is synthetic and command-level:

  • DiscPlaylistTests loads an extended M3U with relative paths and #EXTINF titles, checks wrapping next/previous selection, and rejects an entry-less playlist.
  • DiscPlaylist_ListsEntriesInOrder checks the CLI text/JSON listing.
  • EjectAndInjectCycleTheVirtualTray checks guest eject (0x06), inject (0x07), the Opera tray/disc/spin/ready bits, and the media-access poll.
  • Run_SwapsDiscsAtTheRequestedFrame runs the CLI with --swap-at 0 --swap-disc and checks the swap log and successful result.
  • A manual Release run also swapped from Road Rash (USA).chd to Space Ace (USA).chd at frame 0; the run completed and reported the new disc hash. A known multi-disc title roundtrip remains.

1.1l Portable mode (2026-11-01)

UX-15 coverage is path-resolution level:

  • PortablePathsTests verifies that portable mode resolves to the application directory, non-portable mode resolves to <LocalApplicationData>/5DO, and an explicit FIVEDO_DATA_DIR override wins.
  • A manual CLI run of settings validate --portable --json reported <CLI output directory>/settings.json, proving the flag is stripped before command dispatch and the default path follows the portable directory.
  • Named profiles and a per-instance write lock remain.

1.1m Input display (2026-11-01)

UX-23 coverage is compile/status level:

  • CoreRunnerTests.Status_ReportsTheAudioDerivedTargetFrameRate also asserts the new SkippedFrames and Player1 fields are populated.
  • PbusDecoderTests already cover the button formatting used by the overlay.
  • The Avalonia overlay is build-validated. Lag detection counts frames without a PBUS DMA read (CoreRunnerTests.LagFrames_..., MadamTests.PbusDma_...), and the selected controller port is covered by InputPlayer_RejectsPortsOutsideOneToSix. A screenshot comparison remains optional future polish.

1.1n ODE eject and playlist editing (2026-11-01)

  • DiscPlaylistTests covers create/add/remove/reorder/save with an extended-M3U roundtrip, empty-playlist rejection, and selection repair after removal.
  • CoreRunnerTests proves a guest 0x06 raises GuestDiscEjected once when the tray opens and not again while it remains open; CdromMultiTrackTests verifies TrayOpen, DiscPresent, and the guest-eject counter across eject/inject.
  • SaveStateTests saves and reloads after a disc swap and confirms the second disc's drive state and hash metadata survive.
  • Manual headless evidence: an M3U for D (USA, Europe) disc 1/disc 2 was swapped at frame 300 in a 600-frame run; the disc hash changed to a52a01c7b6d2c8740651ac952b291c0a, 38.4 M instructions executed, and the run completed. This is a timed swap, not yet a game-requested swap.

1.1o PAL/NTSC region modes (2026-11-01)

  • VideoRegionTests verifies NTSC/PAL1/PAL2 geometry and clocks, PAL Quartz wrapping, VDLP start-line mapping, 250,000-cycle PAL fields with 288 visible lines, PAL2 width, and the state region warning.
  • Battle Chess (Europe) with panafz10e-anvil.bin: 600 frames, region=Pal1, fieldRate=50, source=320x288, frame hash 0x88883E88. The captured PNG shows the 3DO logo at the correct PAL size.
  • --bios-auto bios --region auto selected Panasonic FZ-10E (Anvil) and PAL1.
  • FiveDoMachine.SetVideoRegion switches NTSC/PAL1 between frames; VideoRegionTests verifies the 250,000-cycle PAL field and the return to the 208,333-cycle NTSC field.
  • A 300-frame --audio wav run produced 264,609 samples = 6.0002 seconds at 44.1 kHz, confirming the 50 Hz audio/frame cadence.
  • Full NTSC suite: 1,473 tests pass with no changed NTSC hashes.

1.1p Region overrides, instance lock, and CD transcripts (2026-11-01)

  • AppSettingsStoreTests.RegionOverrides_* verifies that a per-game override wins over both the automatic and global region and persists through the settings file.
  • VideoRegionTests.StateSwitchesTheVideoRegionWhenTheStateDiffers saves under PAL1, loads into an NTSC machine, and checks the switch, warning, 250,000-cycle field, and 288-line output.
  • InstanceLockTests acquires the lock twice, checks conflict ownership data, recovers after disposal, creates a missing directory, and reports an unwritable destination.
  • CdromTranscriptTests covers command/status/poll capture, eject/inject, reset replay, and detach/clear. CdromDeviceTests.UnsupportedCommandsPublishAnIllegalRequestStatus checks the 0x17 READ ERROR payload, and ResetCommandClearsTheTransfer... restores the 2048-byte sector size after MODE SET.
  • PrefetchingDiscSourceTests.ReadTrackSector_CachesAndPrefetchesAudioSectors proves a following audio sector comes from the native-sector cache without a second inner read.
  • Manual evidence: Battle Chess (Europe) with panafz10e-anvil.bin, PAL1, 300 frames, --cd-transcript artifacts/cd-transcript.txt produced 393 entries. The first entries are READ_ID, READ_ERROR, MODE_SET, READ_DISC_INFO, READ_TOC, and READ_DATA; the frame hash 0x88883E88 is unchanged from the pre-transcript run.

1.1q Audio device and buffer latency (2026-11-01)

  • AudioBufferConfigurationTests checks clamping at 1/20/100/500/10,000 ms, monotonic queue growth, and that a 100 ms target yields an 80-120 ms queued window.
  • AppSettingsStoreTests.AudioDeviceAndBuffer_RoundtripAndClamp covers device-name trimming, buffer persistence, and 20/500 ms clamping.
  • Manual evidence: 5do audio devices --json listed OpenAL Soft on Headphones (JBL DUET NC) and OpenAL Soft on PG34WQ15R2B (NVIDIA High Definition Audio). A 60-frame PAL Battle Chess run with --audio openal --audio-buffer 20 --audio-device "OpenAL Soft on Headphones (JBL DUET NC)" completed with ok: true, 52,921 samples, and frame hash 0x88883E88.

1.1r Diagnostics bundle (2026-11-01)

  • DiagnosticsBundleTests.Create_WritesScrubbedSettingsLogsAndManifest builds a settings file and log containing the home path and account name, then verifies the zip entries, removed settings keys, placeholder scrubbing, and manifest fields.
  • Create_WorksWithoutOptionalEvidence covers a minimal bundle.
  • Manual evidence: 5do diagnostics bundle --out artifacts/diag-test.zip --json wrote environment.txt, sanitized settings.json, five logs, and manifest.json with 22 scrub replacements; the archive contained no DiscPath or raw user path.

1.1s Per-game input devices (2026-11-01)

  • AppSettingsStoreTests.InputDeviceOverrides_WinOverGlobalAndSurviveRoundtrip checks override precedence, invalid-device normalization, settings roundtrip, and clearing the override back to the global list.
  • AutoInputDevices_SeedKnownLightgunAndTrackballTitles verifies Crime Patrol and Orbatak seeding, idempotence, unknown titles, and the AutoInputDevices switch.
  • The local Crime Patrol (USA) and Orbatak (USA) names match the prefix table; the auto override is written on first disc open and shown in the OSD.

1.1t Save-state format v3 (2026-11-01)

  • SaveState_WritesVersionThreeFeatureFlagsAndSectionTable checks the v3 version, all twelve sections, and the Options/Machine entries.
  • LoadState_AcceptsVersionTwoStates and LoadState_AcceptsVersionOneStates rewrite a v3 header as v2/v1 through StateTestHelper.DowngradeToVersion and confirm a deterministic load.
  • LoadState_UnknownSectionPolicyWarnsOrRejects appends a declared unknown section and checks both Warn and Reject.
  • StateInspectorTests covers v1, v2, and v3 headers, including the uncompressed v1 path; FrameSkipStateTests still proves skip presentation changes only the Machine section.

1.1u State diff (2026-11-01)

  • StateDiffTests.Compare_IdenticalStatesReportsNoChanges and Compare_DifferentFramesReportsChangedSectionsAndOffsets cover equal and diverging states, CPU hash/offset changes, and changed-section counts.
  • Compare_ReportsAnExtraSectionInOnlyOneState appends a declared unknown section and checks the only in b shape.
  • Manual evidence: Battle Chess (Europe) PAL1 states at frames 100 and 120 produced a CPU first difference at byte 4 (39 differing bytes) and a Memory first difference at byte 24,603 (4,184 differing bytes); both report format v3 and the same disc hash.

1.1v Screenshot gallery (2026-11-01)

  • ScreenshotLibraryTests.Parse_ReadsGameStampFrameAndNativeSuffix covers a normal and native companion name.
  • List_SortsNewestFirstAndKeepsUnparsedNames verifies metadata ordering and fallback for a manually named PNG.
  • Delete_RemovesTheFileAndListHandlesMissingDirectories covers delete and an absent directory.
  • The gallery itself is build-validated; thumbnail decoding uses Bitmap.DecodeToWidth and unreadable PNGs fall back to a blank tile.

1.1w GIF recording (2026-11-01)

  • GifRecorderTests covers GIF89a header/trailer, 60-to-30 frame decimation, target-rate clamping, frame limits, size-change skipping, and an empty GIF.
  • Manual evidence: a 40-frame PAL Battle Chess run with --record-gif artifacts/test-recording.gif --record-fps 25 produced 20 frames at 320x288. Pillow loaded every frame and the extracted pixel at (160,120) was a valid RGB color.
  • The encoder uses a 3-3-2 global palette and a NETSCAPE2.0 loop extension; LZW output was validated against Pillow after the code-size transition was corrected.

1.1x Desktop WAV recording (2026-11-01)

  • TeeAudioSinkTests verifies sample forwarding to both destinations, reset forwarding, and null-destination safety.
  • The desktop path reuses WavAudioSink, whose WAV header/data are already covered by AudioAnalyzerTests and AudioPipelineTests.
  • Manual smoke: start File > Record Audio..., run a game, stop, and analyze the WAV with 5do audio analyze (same writer and sample clock as the CLI path).

1.1y Per-game profile editor (2026-11-01)

  • GameProfileJson_RoundtripsAndNormalizes covers export/import and clamping of invalid clock/volume values.
  • CreateEffectiveForGame_AppliesProfileWithoutChangingLiveSettings verifies the machine-construction copy and that global settings are untouched.
  • Run_GameProfileOption_AppliesTimingAndRegion runs the CLI with a PAL1/Arm60/X8/1.5x profile and checks region, armTiming, clioTimers, cdSpeed, and cpuClock in --json.
  • The desktop editor is build-validated; its Save/Import/Export paths use the same GameProfileJson and settings APIs.

1.1z Parser fuzzing (2026-11-01)

  • ParserFuzzTests.StateLoader_MutatedStatesFailWithTypedErrors performs 80 seeded mutations and accepts only InvalidStateException (or a successful benign load).
  • IsoAndChdOpeners_RejectRandomHeadersWithoutCrashing feeds 20 random 8 KB images to each container opener.
  • CueParser_MutatedTextNeverEscapesTypedErrors mutates a valid CUE 80 times and requires successful parses to expose a non-negative sector count.
  • The suite runs in normal dotnet test (~8 s) and is deterministic.

1.1aa Movie frame editor and real-state determinism (2026-11-01)

  • MovieEditorTests covers per-button edits, six-controller replacement, insert/remove ordering and header counts, completion-evidence clearing, and range checks.
  • Real save-state determinism after format v3: Battle Chess (Europe) PAL1 saved at frame 300 with the disc attached and ran 300 more frames to hash 0x88883E88, matching the uninterrupted 600-frame run. Gex saved at frame 300 and continued to 0x945B9A20, also matching.
  • The movie editor window is build-validated; its Load/Save path uses MovieFile, whose container roundtrip is covered by MovieFileTests.

1.1ab Game metadata editor (2026-11-01)

  • MetadataStore_UpsertReplacesByGameIdOrPathAndRemoveClearsIt verifies case-insensitive game-ID replacement, path-key replacement, combined record coalescing, removal, and a second remove returning false.
  • The editor window is build-validated; it uses the same catalog lookup and store calls.

1.1ac Profile selector (2026-11-01)

  • PortablePathsTests.DefaultProfile_IsStoredAndClearedInTheBaseDirectory covers the normalized default-profile.txt roundtrip and removal.
  • The Settings > Paths controls are build-validated; they use AppSettingsStore for create/import/export and PortablePaths.SetDefaultProfile for the next-launch choice.
  • Existing portable, explicit-override, named-profile, and lock tests remain green.

1.1ad Disc integrity verifier (2026-11-01)

  • DiscVerifierTests covers a clean disc and a stable hash, a mid-disc read failure that continues to the end, and a tracked data/audio disc using native sector sizes.
  • Manual evidence: disc verify on Battle Chess (Europe) read 143,810/143,810 sectors, 338,241,120 bytes, hash 0x9147e756840bc269, with no errors.

1.1ae Extended-boot compatibility retry (2026-11-01)

  • tools/compat-sweep.ps1 gained -Files (explicit disc paths) and -RetryFrames (rerun a blank row at a deeper frame count and keep the rendered result).
  • The 18 non-rendered rows from the 3,000-frame sweep were rerun:
    • All five former Not Loadable rows (Crime Patrol, Mad Dog II, Mirai Shounen Conan, Orbatak, Shootout at Old Tucson RE3) complete and render at 20,000 frames with panafz10.
    • Ishida Yoshio Kudan no Igo Seiha, both Konpeki no Kantai discs, all three Myst discs, Olympic Summer Games, and all three Slayer discs render at 20,000 frames.
    • Powers Kingdom (Europe) renders its live-action intro at 65,000 frames with panafz10e-anvil (3,404 colors, 86,442 non-black pixels).
    • WorldCup Super Stadium renders its "Story of the World Cup" screen with panafz10j plus the FZ-10JA Kanji font ROM as BIOS2 (2,425 colors at 20,000 frames).
    • Crash 'n Burn (Japan) remains black or on the Anvil boot screen at 65,000 frames with every local BIOS/font pairing.
  • BiosLibrary now recognizes 1 MB Kanji font images by their 3DO volume-header signature and classifies them as secondary ROMs; the FZ-1J, FZ-1J (no RSA), FZ-10J, and FZ-1J Kanji font hashes are catalogued, and secondary selection prefers the matching FZ-1/FZ-10 family.
  • Final loadability table: 594 discs, 593 Loadable/Rendered, 1 Loadable/No video (known boot failure), 0 Not Loadable. Evidence is in artifacts/compat-20k-10/, artifacts/compat-known/, and the merged artifacts/compat-full/summary-merged.csv.
  • No emulator regression: the full suite remains 1,473 tests, build 0 warnings/0 errors, format clean, inspectcode 0 results.

1.1af Overscan crop (2026-11-01)

  • CoreTypesTests.VdlFrame_CopyCroppedFromShiftsPixelsAndMetadata verifies the pixel shift, palette/background/output-control/DMA metadata, and bounds checks.
  • AppSettingsStoreTests.Overscan_RoundtripsAndClamps covers persistence and the 0/32 px clamp.
  • The desktop path applies the crop before OpenGL/Vulkan/software presentation and in SaveNativeScreenshot; it is presentation-only, so frame hashes and save states do not change. The full suite is 1,473 tests.

1.1ai Screenshot gallery filters, compare, export, and sidecars (2026-11-02)

  • ScreenshotLibraryTests.List_FiltersByGameAndCaptureTime covers case-insensitive game text, inclusive time bounds, native companions, and distinct game names.
  • ScreenshotLibraryTests.Export_CopiesToADestinationDirectory covers directory creation, content copy, and a missing source.
  • ScreenshotMetadataTests covers capture/save/load of renderer, filter, aspect, region, high-resolution, crop, border, profile, and CPU-clock fields; native-companion fallback; invalid JSON; and missing sidecars.
  • The frontend --smoke-test constructs the updated gallery and the new compare window. 1,533 tests pass.

1.1ah Per-edge overscan and void border (2026-11-02)

  • CropInsetsTests covers uniform clamping, the 16-pixel visible-size rule, legacy OverscanPixels fallback, per-edge precedence, SetOverscan clearing the legacy value, and #RRGGBB/#AARRGGBB validation.
  • AppSettingsStoreTests.PerEdgeOverscanAndVoidBorder_RoundtripAndClamp verifies the four edges, border 0-64 clamp, invalid-color fallback, and persistence.
  • AppSettingsStoreTests.GameProfile_FromSettingsCapturesPerGameOverrides and CreateEffectiveForGame_AppliesProfileWithoutChangingLiveSettings verify per-game crop capture/application and live-settings isolation.
  • The frontend --smoke-test constructs the updated Settings window. The crop is presentation-only; CoreTypesTests.VdlFrame_CopyCroppedFromShiftsPixelsAndMetadata keeps the frame-copy contract covered. 1,529 tests pass.

1.1ag 4DO/Opera fix parity (2026-11-01)

  • FixParity.md maps every 4DO FIX_BIT_*: timing 1/2/3/5 are active; timing-6 (Alone in the Dark) and GraphicsStepY (Samurai Shodown) are retired in favor of the generic CEL origin/literal-row model.
  • Local 3,000-frame evidence: DinoPark changes its title scene with the scale-Y fix (0xA1F3E800 vs 0xF2607174); Microcosm (0x4E6A4651) and Super Street Fighter II Turbo (0x1FFBE3FA) are unchanged with/without their bits; forcing timing-6 on Alone in the Dark (0x175B8FE4) or GraphicsStepY on Samurai Shodown (0x86C04E50) is byte-identical.
  • GameCompatibilityTests covers every mapped ID, case-insensitive/full-MD5 lookups, unknown IDs, collisions, and explicit overrides. 1,473 tests pass.

1.1ah Redump DAT matching (2026-11-01)

  • DiscHealthTests parses a synthetic Redump DAT, checks hash normalization, matches a single track by MD5, reports a changed track as a mismatch, and matches two tracks.
  • disc verify artifacts/dat-test.iso --dat artifacts/dat-test.dat --json returned knownDump: true, game DAT Test Game, CRC-32 013C6AE4, MD5 3224F6D33D1FB002DEAE06DD6BEBA153, SHA-1 4EA34675DFFF1D66C21EC9052B387A3C82AAF7BE. A one-byte change produced knownDump: false and exit 1.
  • The full suite is 1,473 tests.

1.1ai Full per-game profile (2026-11-01)

  • AppSettingsStoreTests.GameProfileJson_RoundtripsAndNormalizes now covers BIOS MD5, fix mask, filter, high-resolution, six input devices, and NVRAM mode/version, including invalid-value normalization.
  • GameProfile_FromSettingsCapturesPerGameOverrides verifies the profile snapshot captures the current per-game BIOS/fix/filter/input overrides and NVRAM choices.
  • CreateEffectiveForGame_AppliesProfileWithoutChangingLiveSettings verifies the effective BIOS/fix/filter/high-res/input/NVRAM values and that the live settings and shared override maps are untouched (the clone is deep for the maps).
  • Run_GameProfileOption_AppliesTimingAndRegion now includes FixMode and HighResolution; the CLI reports fixMode: 16 and a 576-line PAL high-res frame.
  • The full suite is 1,473 tests.

1.1aj Interrupt, DMA, and DSP telemetry (2026-11-02)

The CORE-05/06/07 increment was checked with the Debug CLI, panafz1.bin, fresh NVRAM, --frames 3000 --audio null, and the disc named in each row.

Title Frame hash FIQ requests FIQ entries Matched Unmatched Coalesced Mean latency Max latency DSP instructions DSP ticks Peak/tick Mean/tick
Gex (USA, Europe) 0x05F38B50 21,183 14,345 13,855 490 7,328 7,429.8 859,234 88,955,550 2,205,094 106 51.7

The Gex frame hash matches the 3,000-frame key recorded in section 1.1b. A second 3,000-frame Gex run with --clio-trace artifacts/clio-gex.txt recorded 4,714 XBUS DMA transfers, each with a paired clio dma start and clio dma-done line; no clio dma-blocked line appears on the default FreeDO path, as expected.

Lemmings (USA) also booted 3,000 frames to the documented 0x9C2E3236 frame hash with 27,845 FIQ requests / 27,764 entries and a 257,935-cycle maximum latency. These runs are smoke evidence only: the latency distribution still needs an Opera-side timeline before CORE-07 can be marked complete.

1.1ak Profile coverage in bench and selftest (2026-11-04)

The CORE-08 increment closes the remaining CLI surfaces. 5do selftest --json now runs the profile-detection, determinism, non-empty frame, state round-trip, and empty-warning checks for every named profile. The run passed 15/15 checks on the Debug build:

Profile Checks State round-trip
compatible 5/5 PASS
accuracy 5/5 PASS
performance 5/5 PASS

5do bench --frames 600 --profile <name> --json reports the derived profile and timing modes. The synthetic program produces the same final frame hash (0x6CDFBDC5) in all three profiles; the deterministic counters differ as expected:

Profile Profile reported ARM timing Cycles/frame Instructions CPU clock
--profile compatible Compatible FreeDoCompatible 208,336 127,084,960 12.5 MHz
--profile accuracy Accuracy Arm60Datasheet 208,336 127,084,960 12.5 MHz
--profile performance Performance FreeDoCompatible 312,512 190,632,320 18.75 MHz

The measured frames-per-second values are host- and build-dependent and are not a gate; use a Release build with benchmarks/FiveDo.Benchmarks for comparative performance. The accuracy profile does not change this synthetic program's instruction count because it does not poll CLIO timers or CEL status; the timing model is nevertheless active and reported. The profile override in the differential harness is covered by MergeOptions_ProfileOverrideWinsOverDefaultsAndTitle, and the CLI round-trip by Bench_ProfileOptionReportsTiming and SelfTest_CoversEveryNamedProfile.

1.1ak1 40-title Compatible vs Accuracy profile A/B (2026-11-04)

The CORE-08 profile A/B used the same 40-title, 300-frame staged-ISO sample and cached Opera captures as the ARM timing A/B. Only --profile changed:

Profile Matched Diverged Errors
compatible 17 23 0
accuracy 17 23 0

Every title had the same status in both runs, so the Accuracy profile neither closed nor regressed a title in this boot sample. The 594-title, 3,000-frame A/B is still required before the profile can become the default, so Accuracy remains opt-in.

1.2 Second pass: renderer-correct gameplay (2026-10-24)

All runs use the Release build, panafz1.bin, and a fresh NVRAM. The 3,000-frame regression hashes are identical to section 1.1; the longer runs changed because the transposition fix corrected the 3D CEL output.

Title Frames Final PC Frame hash Observation
Road Rash 30,000 0x000681F0 0xEB0AD59C Race renders correctly; DSP running; no allocator error
Road Rash 60,000 0x00095E84 0xBF01DF59 Attract race/FMV loop still alive; DSP running
Road Rash (state, 1,200 frames) 31,200 0x00096300 0x274021FE Control continuation of the 30k state
Wolfenstein 3D 12,000 0x00074B6C 0xC97CA7A5 FLOOR 1-1, SCORE 2800, ITEM 01
Alone in the Dark 16,000 0x0008C608 0x3D257B5C Upright character at the staircase
Lemmings 18,000 0x000205F8 0x5625A4CD Title menu (Play/New Game)
Space Ace 12,000 0x0007D114 0xEF99FED7 Animated demo, no runaway

The exact hashes should be re-read from the --json output rather than copied from memory; the stable observations are:

  • Road Rash: the 3DO controls are B accelerate, A brake, D-pad left/right steer, C attack, and P pause. From the live 26,000-frame state the attract race auto-drives and eventually reaches the WRECK FMV; holding B starts a player race from the GAME LOADING screen. A hard left around frame 31,000 clears the first oncoming taxi. --autopilot road holds B and follows the gray road centre, but it does not yet avoid traffic or recover from a wreck. The race is visually correct and matches the Opera attract-mode road/rider composition.
  • Wolfenstein: from the 12,000-frame demo state, A opens the Select Previews menu; D-pad/A navigate to Select a Level and start a fresh game at FLOOR 1-1 with SCORE 0000000. A live 60-frame state saved after that (0xC9989182) roundtrips: two independent 120-frame continuations both end at 0x370158AC, and a zero-frame load reproduces 0xC9989182 exactly.
  • Alone: the 16,000-frame state has the character upright at the staircase. The control run walks through the room; the scripted D-pad/A run reaches the Begin/Resume main menu.
  • Lemmings: the 18,000-frame state is the title screen. Down + A opens the passcode screen ("Enter the access code"). The keyboard is a 9-column grid with five visible columns; A types the selected letter, B confirms early, and ten letters auto-confirm. A 15-frame D-pad hold moves the selection one key. A wrong code shows "Incorrect" and returns to the title. The title's A/P/X/Left/Right do not leave the attract loop, even over 20,000 further frames. The remaining task is a deterministic path to a valid code such as IJJLDNCCCN (Fun 2).
  • Space Ace: the demo renders animated FMV frames past 12,000 frames.
  • Wolfenstein fresh-game load (open issue): starting a new game from the preview menu reaches a static level screen at 14,000 frames, but it is still there at 54,000 frames. The CD trace shows one 24-sector read every ~33 frames (~87 KB/s) and revisits the 05:11-05:3b region. The attract-mode gameplay at 12,000 frames is unaffected. The Opera-style read-completion status and the 2X drive status bit were tried and reverted: they did not fix the load and changed the golden CD status.
  • Audio (600 frames from the live states): Road Rash 44.1 kHz stereo, 10.03 s, peak 32768, RMS 9051/9146, 191 clipped samples; Wolfenstein 44.1 kHz stereo, 10.02 s, peak 30743, RMS 5620/5457, no clipping. Both are ~737 samples/frame.

Where this leaves us: all five titles boot, render, and produce non-silent DSP audio; four of them have been driven through menus and into gameplay. The remaining gameplay milestone is a scripted race/level completion, plus the Lemmings passcode confirm and an Alone movement/inventory script. See Status.md and ImplementationPlan.md.

1.3 Third pass: gameplay loop and high-resolution rendering (2026-10-24)

All runs use the Release build, panafz1.bin, and a fresh NVRAM.

Wolfenstein 3D gameplay loop. From w3d-fix12k.state (the 12,000-frame demo), w3d-play.txt opens the preview menu and starts FLOOR 1-1 at 14,000 frames. From that state, w3d-walk.txt walks, turns, and fires through 18,000 frames:

14,000 frames: PC 0x0007824C, hash 0x07586A88 (FLOOR 1-1, SCORE 0000000)
18,000 frames: PC 0x00074A04, hash 0xF568EBCA (corridor, guard, pistol)
19,200 frames: PC 0x0001186C, hash 0x74FBBF0C (after walk/turn/fire script)

The HUD, first-person view, guard sprite, and weapon render, and the player moves. The exact values should be re-read from --json; the stable observations are the FLOOR 1-1/score HUD and the visible movement.

Save/load at 18,000 frames:

  • Live state saved at 18,000: frame hash 0xF568EBCA.
  • Zero-frame load: 0xF568EBCA (pixel-identical).
  • Two independent 120-frame continuations: both PC 0x000782B8, hash 0xF568EBCA (the level is waiting for input; the static hash is expected).

Audio from the 18,000-frame gameplay state, 4,000 frames captured with --audio wav: 66.86 s, 44.1 kHz stereo, peak 32767, RMS 2980, 18 clipped samples.

Lemmings high-resolution title. --high-res renders the frontend's 640x480 high-resolution mode:

Low-res  (18,100 frames): 320x240, PC 0x00012B48, hash 0x5CB3039D
High-res (18,100 frames): 640x480, PC 0x00012B48, hash 0xE79E18C5

The high-res capture shows all four title buttons (Play, New Game, Load, Options), the complete logo, and the marquee; it matches the 640x480 reference screenshot. The mode is opt-in because 320x240 titles must not be upscaled. The passcode screen is a 320-wide UI scaled 2x; its five visible keyboard columns are the game's own window, not a framebuffer crop. The five 3,000-frame low-res baseline hashes are unchanged: 0xF4BDF79B, 0x932C7722, 0xEB31636F, 0xA2B49CF0, and 0xC21BB6A1.

Diagnostic input and RAM patches (2026-10-24 late pass)

--poke ADDR=HEX writes bytes into machine RAM after the state loads. It is a diagnostic for input/state investigations, not a game fix. Multiple patches are separated by ;.

# Patch the Lemmings passcode scratch buffer and inspect the next frame
dotnet run -c Release --project src/FiveDo.Cli -- run `
  "$env:FIVEDO_GAMES\Lemmings (USA).chd" --bios bios/panafz1.bin `
  --state-in artifacts/s2/lem-pass.state --frames 3 `
  --poke "0x96574=43414A4A4C444D424556" --frame-out artifacts/poke.png

--autopilot road is an experimental Road Rash driver. It holds B and steers toward the gray road centre (target screen X 230). It follows the road through the first section but does not avoid traffic or recover from a wreck yet.

dotnet run -c Release --project src/FiveDo.Cli -- run `
  "$env:FIVEDO_GAMES\Road Rash (USA).chd" --bios bios/panafz1.bin `
  --state-in artifacts/s2/rr8-30k.state --frames 10000 `
  --autopilot road --capture-dir artifacts/rr-auto --capture-every 1000 `
  --progress 1000 --json

The autopilot/steering tests are in tests/FiveDo.Tests/Host/RoadRashAutopilotTests.cs; the RAM-patch parser tests are in tests/FiveDo.Tests/Host/MemoryPatchTests.cs.

1.4 Fourth pass: Lemmings Fun4 gameplay and a traffic-aware Road Rash driver (2026-10-24)

All runs use the Release build, panafz1.bin, and a fresh NVRAM. Tests are at 572 passing (563 + closing-side, sharp-bend, pacing, coast-floor, and detector-upset cases); dotnet format --verify-no-changes is clean.

Lemmings Fun4 gameplay loop. The passcode screen is reachable from lem-new-end.state (Down + A, see lem-pass-open.txt); the investigation base lem-pass.state is at 21,200 frames (PC 0x20628, hash 0x4321A4A5) with selection on M and one M already typed by the opening A press. Right held 10 frames moves M -> N:

21,200 frames: PC 0x20628, hash 0x4321A4A5 (passcode, selection M, typed M)
21,240 frames: PC 0x230A4, hash 0x73B0DEAD (selection N, typed M, preview N)

lem-toN.txt (21200 Right / 21210 None) reproduces the second line and saves lem-pass-N.state. The 10-letter entry buffer is ASCII but permuted across RAM: display order maps to 0x96577, 0x96576, 0x96575, 0x96574, 0x9657B, 0x9657A, 0x96579, 0x96578, 0x9657F, 0x9657E, with a JC prefix at 0x96572-73. 0x96576 is the blinking preview slot and is refreshed from the current selection every frame, so it must already hold the code's 2nd letter when B is pressed. From the N-selected state, poking the other nine bytes of HNLHCIOECW (Fun 4, 3DO list) and pressing B is accepted:

# From lem-pass-N.state (selection N): poke Fun4 (2nd letter N matches), confirm
dotnet run -c Release --project src/FiveDo.Cli -- run `
  "$env:FIVEDO_GAMES\Lemmings (USA).chd" --bios bios/panafz1.bin `
  --state-in artifacts/s2/lem-pass-N.state --frames 400 `
  --poke "0x96574=48;0x96575=4C;0x96577=48;0x96578=45;0x96579=4F;0x9657A=49;0x9657B=43;0x9657E=57;0x9657F=43" `
  --input-script artifacts/s2/lem-code-HN.txt `
  --frame-out artifacts/s2/lem-fun4-gameplay.png `
  --state-out artifacts/s2/lem-fun4.state --json

lem-code-HN.txt presses B (confirm), A (briefing), A (start). The game shows Code for 1 / Rating = Fun (21,320 frames, PC 0x12B74, hash 0xEA7E3E8D), then the briefing (Press Any Button, 21,440 frames, PC 0x59DD4, hash 0x64C5DEA5), then gameplay (21,640 frames, PC 0x83FEC, hash 0x18F38B4D) with pillars, green ground, lemmings, and the skill HUD.

Save/load at 21,640 frames:

  • Live state saved at 21,640: frame hash 0x18F38B4D.
  • Zero-frame load: 0x18F38B4D (pixel-identical).
  • Two independent 120-frame continuations: both PC 0x83FD8, hash 0x8D628FBD (deterministic; the level animates without input).

Typing ten Ms via A still auto-confirms and correctly reports Incorrect (MMMMMMMMMM is not a valid code), proving the confirm path works with and without pokes. A sequential poke at 0x96574 alone is NOT sufficient: the validation reads the permuted slots, so a sequential poke renders scrambled and reports Incorrect.

Road Rash driver (v8). --autopilot road now holds B, steers to the gray road centre (target X 230) measured in the lower band (rows 190-214, so building facades cannot outvote the asphalt), and adds: lookahead obstacle steering (rows 120-150, steers to the clearer side when the centre is blocked), wall avoidance (does not steer into a walled side when the other side is open), emergency braking (A, no B) when the road ahead is fully blocked, coasting (no throttle) through narrow walled canyons, and stuck recovery (stable off-centre road for 300 frames steers toward the side with visibly more road for 90 frames). Progress lines and run --json report autopilotObstacle and autopilotRecovery alongside autopilotLineX/autopilotSteering.

Best run from rr8-30k.state (30,000 frames) with 10,000 autopilot frames ends at 40,000 frames (PC 0x12850, hash 0x1B9CCE1D). The driver visibly corrects away from the first city wall (obst=1 at 31,000 frames) and triggers stuck recovery (RECOVERY at 33,000-34,000 and 37,000-39,000 frames), but the bike still pins against a city wall around 34,000 frames (speed 0, lineX frozen at ~239, RECOVERY cycling without movement; sustained B+steer/A inputs from the pinned state all converge to the same hash, i.e. the pin is terminal and input has no effect once wedged). The 5.5-mile city race is therefore NOT finished yet; the remaining work is earlier wall anticipation (the lower band still locks onto sidewalks/shadows before the lookahead sees the wall) and/or a track-specific racing line through the 32,000-34,000-frame city section.

Road Rash driver v9: the "pin" was the BUSTED screen (2026-10-25)

The terminal pin above is the race-over BUSTED screen, not a physics wedge. From rr-auto-v8.state (40,000 frames), 400 frames with no input followed by B produce the plainclothes-cop BUSTED screen (hash 0x7D470158); the same happens from rr-v10full.state and rr-v17-40k.state (0x9F44FAA2). No input resumes a lost race, so the earlier input probes could only converge.

Driver v9, covered by tests (572 passing):

  • Road pixels are capped at brightness 120 (was 150), so bright concrete sidewalks and plazas no longer count as asphalt in the lower band.
  • A frame-to-frame lineX jump over 25 px is treated as a detector re-lock: the driver holds straight for 30 frames instead of chasing the false lock. The test is frame-to-frame, so a frozen line cannot latch the hold.
  • Closing-side steering crosses away when the steering-side lookahead fraction is below 0.6 while the other side is above 0.8, coasting during the crossing.
  • Sharp bends (centre below 0.35, one side above 0.8, the other below 0.4) are braked into while steering to the open side.
  • Built-up sections duty-cycle the throttle 40/60 frames, but pacing, caution, and bend lift may never coast more than 20 consecutive frames; the throttle is forced back on because a stopped bike is caught and busted.
  • Stuck recovery triggers after 150 stable off-centre frames (was 300), brakes 30 frames, then re-applies throttle with the recovery steer.
  • Progress lines add la=<left>/<centre>/<right> plus PACE/UPSET; run --json adds autopilotPacing, autopilotUpset, and autopilotLookahead.
dotnet run -c Release --project src/FiveDo.Cli -- run `
  "$env:FIVEDO_GAMES\Road Rash (USA).chd" --bios bios\panafz1.bin `
  --state-in artifacts\s2\rr8-30k.state --frames 10000 --autopilot road `
  --progress 1000 --frame-out artifacts\s2\rr-v17-40k.png `
  --state-out artifacts\s2\rr-v17-40k.state --json

Best v9 run: 40,000 frames, PC 0x000125E4, hash 0xE44B073D. The bike keeps live road detection to about 36,500 frames (the v8 run lost it at ~34,000), then leaves the road and the race is lost. Roundtrip: zero-frame load reproduces 0xE44B073D; two 120-frame autopilot continuations both end at PC 0x000125E4 / 0xE44B073D; pressing a button advances to BUSTED (0x9F44FAA2). The 5.5-mile city race is still NOT finished; the next work is vision/control quality (lane tracking through bends, police/traffic avoidance), not pin recovery.

1.4a Seventh pass: per-game fixes and shell features (2026-10-25)

GameCompatibility now applies the 4DO fix mapping automatically. Alone in the Dark (9f7d72bccaacfc84f230c8d1a7189329) runs with FIX_BIT_TIMING_6 (0x20); its 3,000-frame hash remains 0xF4BDF79B, so the fix does not change the boot screen. The other four corpus IDs have no 4DO fix bits. The CLI reports gameId and fixMode in run --json and info, and accepts --fix-mode 0xMASK / --no-game-fixes. The desktop shell gained the game library, settings dialog, six-player binding capture, numbered save slots, PNG screenshots, and the FPS overlay; 639 tests pass. The Input dialog edits all six players.

1.4b Eighth pass: renderer correctness and desktop validation (2026-10-25)

The five 3,000-frame low-res regression hashes are unchanged after the eighth-pass renderer and desktop work:

Title Final PC Frame hash
Alone in the Dark 0x000C4CB4 0xF4BDF79B
Lemmings 0x0012B3C8 0x932C7722
Road Rash 0x000B9F84 0xEB31636F
Space Ace 0x00076418 0xA2B49CF0
Wolfenstein 3D 0x000699E8 0xC21BB6A1

Commands:

dotnet run -c Release --project src/FiveDo.Cli -- run `
  "$env:FIVEDO_GAMES\Lemmings (USA).chd" --bios bios/panafz1.bin `
  --frames 3000 --json

Desktop renderer validation (Windows, OpenGL 4.0, 1500x978 output):

  • Loaded artifacts/s2/lem-fun4.state through the shell, ran at 60 fps, and captured PNG screenshots through the OpenGL readback path.
  • Sharp-bilinear and xBR screenshots show the Lemmings level with correct colors; before the palette fix the same frames appeared grayscale.
  • The xBR screenshot is edge-smoothed and retains the game image. It is an opt-in filter and is not part of the regression hash.
  • The in-app Logs window showed the renderer-ready and shell-start events; the first-run guide, save-slot thumbnail, and error-dialog paths were exercised manually.

The screenshots are local artifacts and are not committed. The 659-test suite covers the gamepad mapper, log filter, shader source invariants, and slot thumbnails.

1.5 Fifth pass: Road Rash v10 reaches 91,000 live frames (2026-10-25)

All runs use the Release build, panafz1.bin, and the local artifacts under artifacts/s2. Tests are at 578 passing at the time of this pass (572 + close-traffic, urgent-preemption, emergency-steer-hold, stopped/wreck restart, moving-scene, and open-road restart cases); the current suite is 605. dotnet format --verify-no-changes is clean. The five 3,000-frame low-res hashes are unchanged:

Alone 0xF4BDF79B  Lemmings 0x932C7722  Road Rash 0xEB31636F
Space Ace 0xA2B49CF0  Wolfenstein 0xC21BB6A1

Driver v10 changes. Dense captures at 33,800-34,000 showed the loss was close traffic, not a wall: another racer filled the lookahead centre (0.43/0.07/0.17) while every v9 escape rule required a 0.6-wide opening. v10 adds:

  • Close-traffic coast/steer: centre < 0.2, one side in [0.3, 0.6), side difference > 0.2 -> steer into the gap and lift. No brake: a 25-frame full brake from speed 43 stopped the bike and got it busted.
  • Urgent preemption only for close traffic; ordinary obstacle steering stays on the 16-frame decision interval. Applying every obstacle bias immediately oversteered into a taxi at ~32,000.
  • Emergency-brake steer hold: with all lookahead fractions < 0.35, keep the last steer and brake instead of chasing a LineX jump into the tree at 33,890.
  • Stopped/wreck restart: a signature over the left scenery (rows 96-156, x 20-140) and the speed readout stays still for 40 frames -> hold B with straight bars for 180 frames, then retrigger. The centre is excluded because the rider animation reset the first signature. Manual probes: B alone remounts (34,500 and 80,000 -> 52 mph); steering keeps the rider running. A cyclic right/left search was tried and drove the 83,000 run into BUSTED (0x9F44FAA2), so it was reverted.

Best run chain (states are local, gitignored artifacts):

rr8-30k.state (30,000)
  + 6,000  -> rr-v28-36000.state (36,000)
  + 44,000 -> rr-v34-80000.state (80,000, PC 0x00098004, hash 0xD3309E17)
  + 3,000  -> rr-v35-83000.state (83,000, PC 0x00091A70, hash 0x442D7D3D)
  + 8,000  -> rr-v36-91000.state (91,000, PC 0x00085F48, hash 0x02ABE78E)

The 30k->36k and 36k->80k segments used earlier v10 revisions (right-steer crash restarts); the final throttle-only code was verified on 80k->83k and 83k->91k and reproduces the 91,000 state exactly:

dotnet run -c Release --project src/FiveDo.Cli -- run `
  "J:\Panasonic 3DO\Road Rash (USA).chd" --bios bios\panafz1.bin `
  --state-in artifacts\s2\rr-v35-83000.state --frames 8000 --autopilot road `
  --frame-out artifacts\s2\rr-v43-91000.png `
  --state-out artifacts\s2\rr-v43-91000.state --json

Save/load at 91,000 frames:

  • Live state: PC 0x00085F48, hash 0x02ABE78E, cov=0.93, lineX=197.9.
  • Zero-frame load: 0x02ABE78E (pixel-identical).
  • Two independent 120-frame autopilot continuations: both PC 0x000973EC, hash 0x3F27B3D6, lineX=197.875.

Wait states fixed by the media-access pulse (ninth pass).

  • The ~42,000 and ~92,000/100,000 stops were a static, input-ignoring wait; the CD trace contained only READ_ID polls. The fix is in the CD device: two consecutive READ_ID commands with no intervening command and at least 50 frames between them set the media-access bit (0x40) in the poll register. Portfolio uses that event to finish the post-race/post-load transition.
  • From the 42,000 state, the fix advances to the BUSTED screen (~43,500); with A presses it reaches the race-select menu ("The City", "Sierra Nevada", "Napa Valley", "The Peninsula") and starts a new race.
  • From the 100,000 state, the fix advances to BUSTED at ~102,000. From there, A pulses at 300-frame intervals pass the cutscenes and race select and start a new race at ~112,500.
  • The five 3,000-frame corpus hashes are unchanged by the fix. The old 91k autopilot state/hash (0x02ABE78E) was a frozen presentation; the current chain reaches a rider-run/BUSTED transition at 91,000 (PC 0x00093AF4, hash 0x70E1ED8E). Do not use the old hash as a regression key.
  • The autopilot now pulses A during long-static, low-road-coverage, or CD-load states (autopilotMenuAdvance in run --json). This gets through the BUSTED/race-select flow, but it still needs work at the new race start: the bike can spawn at the right edge and stay there with the throttle held. That is a driving/start-line problem, not the CD wait.

Telemetry. Progress lines add cov=<lower-band road fraction> and RESTART; run --json adds autopilotRoadCoverage, autopilotCrashRestart, and autopilotStaticFrames alongside the v9 fields. The ninth pass also reports cdStatus and cdPoll in run --json; a READ_ID-only wait has cdStatus=0xE1 and cdPoll=0x0C.

1.5a Ninth pass: per-title save-state, screenshot, and WAV evidence (2026-10-25)

All runs use panafz1.bin and the Release build. Each case loads a live state, captures the zero-frame frame hash, runs 120 frames, and captures a PNG. The states are local, gitignored artifacts under artifacts/s2 and artifacts/.

Title State Zero-frame PC/hash 120-frame PC/hash
Wolfenstein 3D w3d_gameplay.state (18,000) PC 0x00074A04, 0xF568EBCA PC 0x000782B8, 0xF568EBCA
Lemmings lem-fun4.state (21,640) PC 0x00083FEC, 0x18F38B4D PC 0x00083FD8, 0x8D628FBD
Alone in the Dark alone-fix16k.state (16,000) PC 0x0008C608, 0x3D257B5C PC 0x00095AC0, 0x2ECCE988
Road Rash rr-v36-91000.state (91,000) PC 0x00085F48, 0x02ABE78E PC 0x000821E0, 0x34BAD64C
Space Ace sa-fix12k.state (12,000) PC 0x0007D114, 0xEF99FED7 PC 0x000284EC, 0xA76D03E4

600-frame WAV captures from the same states:

Title Duration Rate Peak RMS L/R Clipped
Wolfenstein 3D 10.03 s 44.1 kHz stereo 18100 2112 / 2229 0
Lemmings 10.04 s 44.1 kHz stereo 22283 4510 / 4526 0
Alone in the Dark 10.06 s 44.1 kHz stereo 20617 1879 / 1926 0
Road Rash 10.03 s 44.1 kHz stereo 32768 9561 / 8787 378
Space Ace 10.04 s 44.1 kHz stereo 4096 1832 / 1832 0

Road Rash still clips in the first 10 s of the 91k state; the game mix is loud there. Space Ace's capture is a mono signal duplicated to both channels. The ~737 samples/frame cadence in this table is the historical FreeDO 8.24 sample clock; see 1.5a2 for the exact 44.1 kHz clock and new captures.

Space Ace interactive path (new). From sa-fix12k.state, X at 12,000 returns to the title/menu (PC 0x0000E12C, hash 0x596B6A81), and A at 12,200 starts the interactive scene. 3,000 frames later the run is at 15,000 (PC 0x000765F0, hash 0x7E780D93). The 15,000 state roundtrips: zero-frame load 0x7E780D93; two independent 120-frame continuations both PC 0x000284EC, hash 0x70B16422. A 600-frame WAV from the gameplay state is 10.04 s, peak 4096, RMS 1832.

Alone in the Dark menu/intro. From alone-fix16k.state, the sequence Up, Left, A, Down reaches the Begin a new game / Resume a saved game menu at 16,380 (PC 0x000125E4, hash 0x22631207). A at 16,600 starts the intro/character scene (PC 0x00095A24 at 17,200, hash 0x9C607707). The intro then stays completely still and ignores every button, with only READ_ID polling in the CD trace; it has the same wait signature as the Road Rash stops.

Road Rash 42k wait. A fresh run from rr8-30k.state reaches 42,000 with PC 0x0001048C, hash 0xCBDF6DDB, lineX=177.8, cov=0.90. Before the media-access fix, B, A, B+Left, B+Right, A+Left, A+Right, and bare D-pad inputs all left the frame hash unchanged and the CD trace contained only READ_ID. See 1.5b: the media-access pulse now advances this screen to BUSTED and the race-select/new-race flow.

1.5a2 Exact DSP sample clock and live-audio validation (2026-10-27)

The DSP sample clock now accumulates the actual ARM cycle delta, producing exactly 44,100 samples per emulated second. This removes the FreeDO 8.24 over-count that made audio ~0.7% sharp and kept the live ring overfilled.

Check Before After Reference
Lemmings, 600-frame boot samples 444,097 441,015 Opera 441,441 at 59.94 Hz
Lemmings, 600-frame boot duration 10.070 s 10.000 s Opera 10.010 s
Road Rash, 6,000-frame WAV samples 4,429,529 4,410,188 100.0 s at 44.1 kHz
Road Rash, spectral peak in 10-20 s 442.1 Hz 442.1 Hz Opera 442.1 Hz

The old and new Road Rash WAVs have the same peak (29,951), RMS within 0.2%, and identical spectral peaks; only the sample count/timebase changed. Alone and Lemmings 3,000-frame screenshots are pixel-identical before/after; Road Rash and Space Ace show adjacent FMV frames; Wolfenstein differs only in the star field.

The Lemmings Fun 4 D-pad script still reaches gameplay at 23,000 frames with the exact clock (PC 0x00083FE8, hash 0x63C02DED, TIME 4:57); the old pre-clock key 0xB7846C2D is historical. The in-game timer differs by a few seconds because the audio clock changed the frame/FMV timing.

600-frame WAV captures from live states with the exact clock:

Title State Duration Samples Peak RMS Clipped Silent
Wolfenstein 3D w3d_gameplay.state (18,000) 10.000 s 441,020 18,100 2,168 0 8
Lemmings lem-dpad-gameplay.state (23,000) 10.000 s 441,020 8,192 1,661 0 27,903
Gex gex-gameplay.state (15,000) 10.000 s 441,019 12,964 2,089 0 6,923
Theme Park tp-buttons.state (15,000) 10.000 s 441,018 0 0 0 441,018
Road Rash rr-t198-92k.state (92,000) 10.000 s 441,019 32,768 11,012 2,036 0

Road Rash's race mix still reaches full scale and clips; that is the game's mix, not the output path. Theme Park's captured menu is silent. audio analyze <wav> --json prints the same peak/RMS/clipping/DC-offset statistics for any capture. A 36,000-frame capture from rr-t198-92k.state produced 26,461,238 samples (600.028 s at 44.1 kHz), only 0.005% above the nominal 36,000 x 735 samples; the exact clock does not drift over ten minutes. That continuation reaches long static wait sections (97.9% silent frames), so it is a cadence check rather than a gameplay-audio sample.

DSP start/stop. Road Rash starts the DSP between 500 and 1,000 frames and keeps it running through 6,000; Lemmings runs it from boot through 25,000. No title in the current corpus was observed stopping the DSP mid-run. Resets and state loads call IAudioSink.Reset, which drops the live ring/queue so a loaded timeline does not replay old samples.

Live audio. OpenAlAudioOutput.TryCreate now brings the OpenAL context current on the audio thread and the build/publish copies the bundled OpenAL Soft 1.23.1 runtime (OpenAL32.dll on Windows). A 300-frame real-time run with the OpenAL null backend completed in 5.000 s with 220,510 samples (735.03/frame), 529 dropped samples, 14 padded buffers, and ~89 ms output latency; the frame pacer followed the 44.1 kHz clock. The trimmed single-file CLI (win-x64 and linux-x64) and the opt-in NativeAOT CLI both open the same backend; the Linux check used the published libopenal.so and ALSOFT_DRIVERS=null under WSL. The opt-in FIVEDO_AUDIO_TESTS=1 xUnit test plays 1 s of a 440 Hz tone through the real device.

90-second Opera comparison (2026-10-28). A 90 s Road Rash boot capture from the exact-clock DSP (artifacts/rr-90s-machine.wav) was compared with a 95 s WASAPI loopback of the Opera reference (artifacts/loopback-opera-90.wav). Both are 44.1 kHz stereo and cover the same boot/attract sequence:

Metric 5DO DSP raw Opera host
Duration 90.0 s 95.0 s
Peak 0.914 0.914
RMS 0.0615 0.0644
20-200 Hz band 52.7% 50.4%
200-2k Hz band 30.4% 33.4%
2k-8k Hz band 16.6% 15.9%
8k-20k Hz band 0.3% 0.3%
400-500 Hz tone 441.4 Hz 441.4 Hz

The log-power spectra are 0.9994 cosine-similar over 20 Hz-20 kHz. The 10 s RMS envelopes correlate at r = 0.9685 after excluding the first window (the captures start a few seconds apart in boot), and the per-window levels match from window 2 onward. This closes the loudness/spectrum comparison. The only remaining audio check is a manual listen for clicks when a game stops/starts the DSP; no current corpus title was observed stopping it mid-run.

1.5b Ninth pass: CD media-access pulse unblocks the post-race wait (2026-10-25)

The wait states at Road Rash ~42k/~92k/100k and Alone's intro share a CD signature: the CD task polls READ_ID (cmd 0x83) every 60 frames with no other commands, the frame is static, and every input is ignored. The ninth pass found that Portfolio expects a media-access event after the drive has been idle. The CD device now sets poll bit 0x40 when two consecutive READ_ID commands arrive with no intervening command and at least 50 frames between them. Boot traffic never has that pattern, so the five 3,000-frame hashes are unchanged.

Evidence:

  • Road Rash rr-42k.state + 3,000 frames with A: BUSTED by ~43,500; final PC 0x00020640, hash 0x5B3159A9 (the BUSTED screen).
  • Road Rash rr-100k-base.state + 3,000 frames with A: BUSTED by ~102,000. From there A every 300 frames passes the cutscenes and race-select menu ("The City" / "Sierra Nevada" / "Napa Valley" / "The Peninsula") and reaches a new race at 113,000 (PC 0x0007DCBC, hash 0x277CCED7).
  • The autopilot now pulses A during long-static, low-road-coverage (coverage below 0.5), frozen-road-line, or CD-load states; run --json reports autopilotMenuAdvance. A fresh 30k run reaches the post-race cutscene and a new race around 70-80k, but still needs start-line handling.
  • Alone's intro uses the same wait signature, but the media pulse alone does not advance it: alone-intro.state + 3,000 frames with no input stays at hash 0x9C607707. A every 300 frames advances the man-at-desk cutscene to the room/mansion/title scenes: 29,200 frames, PC 0x00020640, hash 0x62304BCF; a further 10,000 frames ends at 39,200 (PC 0x0008BCA4, hash 0xD1FABFD3). A blind A pulse is enough to leave the wait but not yet enough to reach a stable gameplay room.
  • Alone movement loop. From alone-intro-A2.state (39,200), the sequence Up, Left, A, Down, Right moves the character around the room (the sprite visibly walks between frames). At 39,800: PC 0x0008D0EC, hash 0x44C7576B. The state roundtrips: zero-frame load 0x44C7576B; two independent 120-frame continuations both PC 0x0008BF00, hash 0xE490C8C7. This is the first scripted Alone movement/save-load loop.
  • Wolfenstein's fresh-game load (w3d-loadlong.state, 54,000 frames) also stays static with no input (0x07586A88). A every 300 frames advances to the character-select screen and FLOOR 1-1 at 64,000 (PC 0x00069F3C, hash 0x2F5B9864). The media pulse alone does not advance it; A is required.
  • Road Rash race/menu cycling. From rr-fix-60k.state, the autopilot now passes the GAME LOADING/Get ready screen and is racing at 80,000 (PC 0x00092FF8, hash 0x1074EF3F). A further 20,000 frames cycle through race -> wreck/menu -> new race several times and end at 100,000 (PC 0x0008214C, hash 0x0D46CF56) with the bike on a sidewalk, not a finished race. The restartStuck rule (release throttle and pulse A after 180 restart frames) is what gets through the start-line screens.

The 91,000-frame autopilot chain now reaches a rider-run/BUSTED transition (PC 0x00093AF4, hash 0x70E1ED8E) instead of the old frozen 0x02ABE78E frame. The old hash is historical, not a regression key. The five 3,000-frame corpus hashes remain the regression gate and are unchanged.

CD disc-end fix (2026-10-23)

CdromDevice.InitCd originally reproduced FreeDO's double-add of the 150-frame lead-in (filesize = sectors + 150 followed by LBA2MSF, which adds 150 again). The CD DIPIR compares the reported disc end with the volume header's block count and rejects the disc when the end exceeds it by more than 512 frames. Lemmings has 302 blocks of padding and passed; Road Rash (450), Space Ace (450), and Wolfenstein (382) were rejected and fell into the shell's InsertDisc path. The device now reports MSF = sectorCount + 150, i.e. the real end. This is an intentional divergence from the FreeDO reference, which still has the bug.

Reference comparison

A local FreeDO harness (References/freedo-master, not committed) was built from the same source revision and run against an ISO dump of the same CHD. Before the disc-end fix, Road Rash's kernel debug (KPRINT) stream was byte-for-byte identical to FreeDO through 30,000 frames, including GetCDType:Got error d556f106, FFS-Severe-System-extended-Volume offline, and the $panasplash/PanaSplash, $insertdisc/InsertDisc, and $eldemo/ELDemo loads. That parity confirmed the failure was in the shared CD layer rather than in the ARM or video port. FreeDO still has the double-add and still fails on these discs; the C# port now intentionally diverges.

After applying the same disc-size correction to the harness host (returning sectorCount - 150), FreeDO also reaches Road Rash's own OS and stops at the audio path with >>>ARM TO DSP FIFO DIRECT!!!. The C# port reaches the slightly later SAudio channel 0 failed; error = -6996 message.

This is not a shared DSP limitation. The 2026-10-23 Opera comparison (below) shows Opera running the same titles past this point. The FreeDO harness result only describes the older FreeDO baseline; it is not the compatibility ceiling.

Opera reference run (2026-10-23)

The data track of each CHD was extracted to ISO with 5do disc dump, then run with the bundled Windows OperaCore (References/CPP_SimpleGui/Cores/OperaCore) and the same panafz1.bin BIOS. The SimpleGui wrapper was launched with --debug-messages -n <iso>; screenshots were captured at 15/35/60 s.

Title Opera result Our port after 2026-10-24 DSP fix
Road Rash Intro FMV rendered for 60 s Intro FMV renders; DSP/audio runs; 30,000-frame run is stable
Wolfenstein 3D Intro rendered for 60 s Boot/OS screens render; DSP/audio runs; 30,000-frame run is stable
Alone in the Dark Intro/loading rendered for 60 s Boot/logo renders; DSP/audio runs; 10,000-frame run is stable
Space Ace Demo intro for 60 s, no runaway Demo intro renders past 10,000 frames; no runaway

Controls:

  • The same ISO now also runs in our port, so the CHD reader is not the cause.
  • Arm60.Ncycle = 2 changed the execution point but was not needed for the fix.
  • The DSP immediate-shift bug was the common cause of SAudio -6996 and the Space Ace null-buffer copy. See Compatibility.md.

The full method, reference hierarchy, and next actions are in ReferenceStrategy.md and ImplementationPlan.md.

The DSP produces roughly 738 samples per frame (44.1 kHz at 60 fps) for all five titles, and WAV capture is non-silent (Road Rash peaks around 8,800; Space Ace around 4,096 in the first 1,200 frames). Audio content still needs a listening check against the real hardware or Opera. The headless core runs a 3,000-frame smoke test in roughly 4-8 seconds on the development machine; that is not a 4K presentation benchmark.

Diagnostics used

The CLI now exposes the tools used for the comparison:

# Kernel debug output and the PC of every character (plus LR/SP/stack words)
dotnet run --project src/FiveDo.Cli -- run "$env:FIVEDO_DISC" `
  --bios "$env:FIVEDO_BIOS" --frames 3000 `
  --kprint artifacts/kprint.txt --kprint-pc artifacts/kprint-pc.txt

# CD command/sector trace
dotnet run --project src/FiveDo.Cli -- run "$env:FIVEDO_DISC" `
  --bios "$env:FIVEDO_BIOS" --frames 3000 --cd-trace artifacts/cd.log

# CLIO timer/FIQ register trace
dotnet run --project src/FiveDo.Cli -- run "$env:FIVEDO_DISC" `
  --bios "$env:FIVEDO_BIOS" --frames 3000 --clio-trace artifacts/clio.log

# Instruction ring buffer around a stop condition
dotnet run --project src/FiveDo.Cli -- trace "$env:FIVEDO_DISC" `
  --bios "$env:FIVEDO_BIOS" --trace-frames 600 --ring 8192 `
  --stop-kprint "Chunk XTRA" --out artifacts/trace.txt

# Low-memory write watchpoint (writer PC, value, size, registers, caller)
dotnet run --project src/FiveDo.Cli -- run "$env:FIVEDO_DISC" `
  --bios "$env:FIVEDO_BIOS" --frames 700 `
  --watch-lowmem artifacts/watch.txt --watch-from-frame 681 `
  --watch-start 0 --watch-end 0x100

# RAM and DSP memory dumps at the end of a run
dotnet run --project src/FiveDo.Cli -- run "$env:FIVEDO_DISC" `
  --bios "$env:FIVEDO_BIOS" --frames 3000 `
  --dump-ram artifacts/ram.bin --dump-ram-offset 0 --dump-ram-length 0x300000 `
  --dump-dsp artifacts/dsp

# DSP instruction ring at a KPRINT stop
dotnet run --project src/FiveDo.Cli -- trace "$env:FIVEDO_DISC" `
  --bios "$env:FIVEDO_BIOS" --frames 0 --trace-frames 3000 `
  --stop-kprint "SAudio channel" --dsp-trace artifacts/dsp-trace.txt `
  --dump-dsp artifacts/dsp-stop --out artifacts/trace.txt

# Scripted controller taps: frame, then a comma-separated button list.
# Buttons are A,B,C,P,X,L,R,Up,Down,Left,Right; None releases all.
@'
# frame buttons
0    None
120  P
130  None
240  A
250  None
'@ | Set-Content artifacts/input.txt
dotnet run --project src/FiveDo.Cli -- run "$env:FIVEDO_DISC" `
  --bios "$env:FIVEDO_BIOS" --frames 3000 `
  --input-script artifacts/input.txt --frame-out artifacts/input.png

# Script frames are absolute. With --state-in, either use the state's frame
# numbers (for example 91000) or add --script-base state to shift a script that
# starts at 0 by the loaded state's frame count.
dotnet run --project src/FiveDo.Cli -- run "$env:FIVEDO_DISC" `
  --bios "$env:FIVEDO_BIOS" --state-in artifacts/live.state --frames 300 `
  --input-script artifacts/input.txt --script-base state --json

# Controller packet trace (one 16-byte PBUS packet per frame)
dotnet run --project src/FiveDo.Cli -- run "$env:FIVEDO_DISC" `
  --bios "$env:FIVEDO_BIOS" --frames 600 `
  --input-script artifacts/input.txt --pbus-trace artifacts/pbus.log

# CEL command/geometry trace
dotnet run --project src/FiveDo.Cli -- run "$env:FIVEDO_DISC" `
  --bios "$env:FIVEDO_BIOS" --frames 600 --cel-trace artifacts/cel.log

# Disassemble raw little-endian ARM words from a RAM dump
dotnet run --project src/FiveDo.Cli -- disasm artifacts/ram.bin `
  --offset 0x12380 --count 100 --base 0x0

1.6 Tenth pass: Lemmings pure D-pad passcode and CRT bloom (2026-10-26)

Lemmings Fun 4 passcode typed with the D-pad. The ninth-pass path used a diagnostic RAM poke for the passcode. The tenth pass types HNLHCIOECW entirely with the controller from a passcode-screen state with an empty entry and the selection on V (artifacts/lem_pass_fresh.state, 21,700 frames). The script is committed as scripts/lem-fun4-dpad.txt; it uses 15-frame direction holds and 5-frame A taps, with one release frame between actions so the CLI's frame coalescing cannot swallow the release.

dotnet run -c Release --project src/FiveDo.Cli -- run artifacts/s2/lem.iso `
  --bios bios/panafz1.bin --state-in artifacts/lem_pass_fresh.state --frames 1300 `
  --input-script docs/scripts/lem-fun4-dpad.txt --script-base state `
  --frame-out artifacts/lem-dpad-gameplay.png `
  --state-out artifacts/lem-dpad-gameplay.state --json

The game shows the Fun 4 / Level 1 / Save the Lemmings briefing at 22,600 (PC 0x00020628, hash 0xC1DF57C5) and starts gameplay after the A tap at 22,650. At 23,000 the run is in the level (PC 0x00083FD8, hash 0xB7846C2D) with the HUD and lemmings visible. Roundtrip:

  • Zero-frame load of lem-dpad-gameplay.state: PC 0x00083FD8, hash 0xB7846C2D (pixel-identical).
  • Two independent 120-frame continuations: both PC 0x000230A4, hash 0xEDEB1B9D.

This closes the "pure D-pad typing of all ten letters is not scripted" item in KnownIssues.md. The script starts at the passcode screen, not at boot; the title-to-passcode transition is still the existing lem-pass-open.txt (Down, A) path.

CRT bloom. RendererSettings.Bloom (0-1, default 0.35) and BloomThreshold (0-1, default 0.6) add a 3x3 native-resolution glow around bright areas in the CRT filter. The settings dialog exposes a CRT bloom slider. Bloom is ignored by the non-CRT filters and by the CPU fallback. The shader is covered by source-invariant tests; hardware GPU timing for the extra samples is still pending.

HQx 2x/3x/4x (historical CPU path). This pass predates the 2026-11-07 GPU rule-table implementation. VideoFilter.Hqx, Hqx3x, and Hqx4x ran Maxim Stepin/Cameron Zemek's hq2x/hq3x/hq4x on the CPU-expanded BGRA frame, uploaded the scaled result, and finished with sharp-bilinear. A 32x24 deterministic reference image matches the C++ byte hashes 0xA400C421 (2x), 0x0C406F97 (3x), and 0xBB17969C (4x); a uniform image is unchanged. The reference's RGBtoYUV table loop stops before the last entry, so pure white is deliberately excluded from the reference image. A 320x240 synthetic benchmark measures about 2.3/4.2/7.5 ms per frame for 2x/3x/4x on checkerboard content and about 12/13/15 ms on random noise, so 3x/4x are opt-in and not recommended with high-res on slow CPUs. The current implementation runs the same rule table on the GPU; see Video.md and Performance.md for the 2026-11-07 numbers.

1.7 Fifteenth pass: additional boot/render/audio corpus (2026-10-26)

Seven additional titles were booted with panafz1.bin, a fresh NVRAM, and 12,000 frames. Each rendered a real frame and produced a deterministic save state. The frames are not committed; the hashes below are the regression keys.

Title File Frames Final PC Frame hash Observation Audio (600 frames)
Gex Gex (USA, Europe).chd 12,000 0x000876E8 0x6B8760C9 3D platform level with HUD 10.03 s, peak 29733, RMS 3029/3145, 0 clipped
Need for Speed Road & Track Presents - The Need for Speed (USA, Europe).chd 12,000 0x0001055C 0x95730015 Car FMV and menu 10.02 s, peak 29951, RMS 5309/5022, 0 clipped
Super Street Fighter II Turbo Super Street Fighter II Turbo (USA).chd 12,000 0x00010548 0x110E7DC5 Attract demo (title, character intro, CPU fight, ending) at 2,000-10,000; black transition at 12,000 10.10 s, peak 21248, RMS 1569/1502, 0 clipped
Total Eclipse Total Eclipse (USA).chd 12,000 0x000106A4 0x280E8346 Spaceship intro/gameplay over terrain 10.06 s, peak 16382, RMS 2760/2760 (mono duplicated), 0 clipped
Killing Time Killing Time (USA) (RE1).chd 12,000 0x000A4128 0xF23DE48C Title menu (START/OPTIONS/LOAD GAME) 10.08 s, peak 32333, RMS 5609/6357, 0 clipped
Theme Park Theme Park (USA).chd 12,000 0x0002063C 0x55B0AC11 Title/menu screen 10.11 s, silent at this menu (peak 0), DSP running
Return Fire Return Fire (USA).chd 12,000 0x000284F4 0xD57E56D7 Menu screen (controller/map panels) 10.12 s, peak 29765, RMS 2600/2600 (mono duplicated), 0 clipped

Commands:

$g = $env:FIVEDO_GAMES
$b = "bios\panafz1.bin"
$t = "Gex (USA, Europe).chd"
dotnet run -c Release --project src/FiveDo.Cli -- run "$g\$t" --bios $b `
  --frames 12000 --frame-out artifacts\gex-12k.png `
  --state-out artifacts\gex-12k.state --json

# Save-state roundtrip: zero-frame load + two 120-frame continuations
dotnet run -c Release --project src/FiveDo.Cli -- run "$g\$t" --bios $b `
  --state-in artifacts\gex-12k.state --frames 0 --json
dotnet run -c Release --project src/FiveDo.Cli -- run "$g\$t" --bios $b `
  --state-in artifacts\gex-12k.state --frames 120 --json

# Audio capture from the 12,000-frame state
dotnet run -c Release --project src/FiveDo.Cli -- run "$g\$t" --bios $b `
  --state-in artifacts\gex-12k.state --frames 600 --audio wav `
  --wav-out artifacts\gex-12k.wav --json

Roundtrips (zero-frame load matches the saved hash; two independent 120-frame continuations match each other):

Title Zero-frame hash 120-frame PC / hash (twice)
Gex 0x6B8760C9 PC 0x00079414, 0xC682B436
Need for Speed 0x95730015 PC 0x0001281C, 0xB1FB873D
Super Street Fighter II Turbo 0x110E7DC5 PC 0x00020648, 0x110E7DC5
Total Eclipse 0x280E8346 PC 0x000205F4, 0xDD148BD5
Killing Time 0xF23DE48C PC 0x000284EC, 0xF23DE48C
Theme Park 0x55B0AC11 PC 0x00019080, 0x55B0AC11
Return Fire 0xD57E56D7 PC 0x0002E21C, 0xD57E56D7

The 4DO game database applies fix mask 0x2 to Super Street Fighter II Turbo; the other six titles use mask 0. Theme Park's menu is silent; this is recorded as an open audio question, not a pass for title music.

2. BIOS images

The local BIOS directory contains several variants. Only 1 MB images are accepted by BiosImage; the 512 KB arcade image and the truncated Kanji image are not valid inputs for the emulator.

File Size (bytes) Intended use
panafz1.bin 1,048,576 Primary NTSC test BIOS (MD5 f47264dd47fe30f73ab3c010015c155b)
panafz10.bin 1,048,576 Alternate NTSC BIOS
panafz10-patched.bin 1,048,576 Patched NTSC BIOS; required by the Arcade CD-ROM titles
goldstar.bin 1,048,576 GoldStar BIOS
sanyotry.bin 1,048,576 Sanyo BIOS
panafz10e-anvil.bin 1,048,576 European/PAL Anvil BIOS (MD5 a48e6746bd7edec0f40cff078f0bb19f); required by Battle Chess (Europe)
panafz10e-anvil-patched.bin 1,048,576 European/PAL Anvil BIOS with the RSA check disabled (MD5 cf11bbb5a16d7af9875cca9de9a15e09)
panafz10ja-anvil-kanji.bin 1,048,576 Japanese Anvil/Kanji BIOS
panafz1-kanji.bin 933,636 Not accepted (not 1 MB)
3do_arcade_saot.bin 524,288 Not accepted (arcade ROM, not a 1 MB BIOS)
panafz10.bin.fstree 17,459 Metadata, not a BIOS

Use panafz1.bin for the default smoke test. Use the Anvil and GoldStar variants only when validating compatibility-specific code paths. panafz10e-anvil.bin is the European/PAL build and is the reference BIOS for European-only discs such as Battle Chess (Europe), which rejects NTSC system software.

3. CLI commands

Set the environment variables used by the opt-in tests. FIVEDO_GAMES is the corpus root and is different on each machine; everything else is derived from it.

$env:FIVEDO_BIOS  = (Resolve-Path "bios\panafz1.bin")
$env:FIVEDO_GAMES = "J:\Panasonic 3DO"        # main collection
# $env:FIVEDO_MAME = "G:\MAME\MAME Software List CHDs\3do"
$env:FIVEDO_DISC  = "$env:FIVEDO_GAMES\Road Rash (USA).chd"
Get-ChildItem $env:FIVEDO_GAMES

Inspect a disc without booting it:

dotnet run --project src/FiveDo.Cli -- info "$env:FIVEDO_DISC" --json

List or extract the 3DO (Opera FS) file system:

dotnet run --project src/FiveDo.Cli -- fs list "$env:FIVEDO_DISC" --json
dotnet run --project src/FiveDo.Cli -- fs extract "$env:FIVEDO_DISC" `
  --out "$env:TEMP\fivedo-fs" --filter "Rash/"

fs uses the first-party FiveDo.Storage.OperaFileSystem parser and accepts ISO, BIN, CUE+BIN, and CHD images; info and disc dump use the emulator's own readers. The Lemmings ISO/CHD pair was cross-checked with 482 identical entries and 461 byte-identical extracted files.

Run a fixed number of frames and dump the last frame:

dotnet run --project src/FiveDo.Cli -- run "$env:FIVEDO_DISC" `
  --bios "$env:FIVEDO_BIOS" `
  --frames 600 `
  --nvram artifacts/nvram.ram `
  --frame-out artifacts/road-rash-600.png

The CLI loads a fresh 32 KB NVRAM volume when --nvram is omitted. Pass --nvram to keep the same save data across runs; games that check save-space (Alone in the Dark) need a valid volume to get past their startup warning.

Capture audio for listening:

dotnet run --project src/FiveDo.Cli -- run "$env:FIVEDO_DISC" `
  --bios "$env:FIVEDO_BIOS" `
  --frames 600 `
  --audio wav `
  --wav-out artifacts/road-rash-600.wav

Run the deterministic self-test, the synthetic benchmark, and the CHD/ISO prefetch comparison:

dotnet run --project src/FiveDo.Cli -- selftest
dotnet run --project src/FiveDo.Cli -- bench --frames 600
dotnet run --project benchmarks/FiveDo.Benchmarks -- --frames 600 `
  --bios bios/panafz1.bin --disc "$env:FIVEDO_GAMES\Lemmings (USA).chd" --json
dotnet run --project benchmarks/FiveDo.Benchmarks -- --frames 600 `
  --bios bios/panafz1.bin --disc "$env:FIVEDO_GAMES\Lemmings (USA).chd" --prefetch --json

Desktop validation:

dotnet run --project src/FiveDo.Frontend.Avalonia

Open the BIOS, open a disc, press Play, and check video, audio, input, pause, reset, quick save/load, screenshots, fullscreen (button or F11/Escape), the Hi-Res restart path, and the About dialog. If OpenGL initialization fails, the status line should report the fallback and the CPU WriteableBitmap path should render.

4. Test matrix

Record the BIOS, disc, and commit for every result. A pass means the expected screen/audio/input behavior was observed; "boots without crashing" alone is not a compatibility pass.

Area Cases
BIOS boot panafz1.bin, panafz10.bin, goldstar.bin, Anvil BIOS
Disc boot Alone in the Dark, Lemmings, Road Rash: BIOS animation -> title/menu
Gameplay Reach gameplay in each title; exercise controller buttons
Audio Title music/SFX and in-game audio; WAV capture is non-silent and correctly pitched
Save states Save in-game, run 60 frames, load, verify frame/state hash and continued play
CHD streaming Boot, seek, and stream from each CHD without frame stalls or read errors
Rendering Accurate/nearest, sharp-bilinear, CRT, integer scaling, fullscreen, window resize
Performance Native, 1080p, and 4K output; record p50/p95/p99 frame time and GPU time
Lifecycle Pause/resume, reset, open/replace disc, close during play, reopen

1.1af Library disc-health verification (2026-11-04)

  • DiscHealthCacheTests covers the persistent cache roundtrip, malformed files, stale size/timestamp records, badge text for every status, summary counts, missing files, unreadable images, DAT matches and mismatches, and progress reporting.
  • CliLibraryVerifyTests runs library verify in-process with a synthetic ISO and DAT: the JSON report says verified, the cache file is written, a corrupt image reports ReadError and exits 1, and an empty folder returns the documented error.
  • Manual evidence: 5do library verify --folder "D:\Samples\3DO" --cache <temp> read all five local CHDs (Alone in the Dark, Lemmings, Road Rash, Space Ace, Wolfenstein 3D; 1.19 GB) in 91 seconds with 0 read errors and wrote a NoDat record for each; no DAT is available locally, so the DAT path is covered by the synthetic tests.

5. Rules

  • Never commit BIOS ROMs, game images, extracted assets, screenshots containing copyrighted content, or WAV captures of copyrighted audio.
  • Keep generated artifacts under artifacts/, which is gitignored.
  • Do not add game-specific hacks without evidence; record the disc, BIOS, and observed behavior in docs/Compatibility.md.
  • Real-content tests stay opt-in. CI must continue to pass with synthetic ROMs and discs when FIVEDO_BIOS/FIVEDO_DISC are unset.

1.1al SYS-07 memory capacity configurations (2026-11-04)

MemoryConfiguration and MemoryLayout cover the eight Opera layouts (21, 22, 41, 42, 81, 82, E2, F1). MemoryConfigurationTests (34 cases) covers the derived geometry, the MADAM RED SYSBITS values, all eight DRAM/VRAM windows, byte/halfword/word decoding at the new boundaries, the stock 3 MB boundary, TryParse, CLI reporting, state round-trip, and a legacy version-2 memory section.

  • Homebrew probe: a synthetic ARM ROM writes 0x11, 0x22, 0x33, and 0x44 to the top of DRAM (0x003F0000), the VRAM base (0x00410000), and the top of the second VRAM bank (0x005F0000), reads them back, and stores them at 0x1000. It also records MADAM RED SYSBITS at 0x2004. With --memory 42 the results are the four values, marker 0xA5 at 0x2000, and SYSBITS 0x32.
  • State: a 4 MB DRAM / 2 MB VRAM machine stores 0xDEADBEEF at 0x00500000; the state reloads that value and the layout into a stock machine, with the State memory capacity ... switched to ... warning.
  • Retail compatibility: Lemmings (USA) boots unchanged with the stock layout (20,000-frame hash 0xE84356D0). The 2 MB DRAM / 2 MB VRAM layout still runs and keeps the DSP active. Layouts with more than 2 MB of DRAM (4/1, 4/2, 8/1, 8/2) stop on the boot screen, which matches Opera's documentation that expanded memory is a homebrew option and can break retail software. The stock layout remains the default and its golden hashes are unchanged.

1.1am SYS-15/SYS-16 eight-port PBUS (2026-11-04)

  • PbusEncoder.MaxPorts is 8; MaxControllers remains 6 for the legacy 4DO packet. InputDeviceSet, PbusSource, InputManager, the chain encoder/decoder, settings, and the CLI accept ports 1-8. Ports 7-8, or any peripheral, select the Opera chain.
  • --active-devices 0-8 and Settings > Emulation > Active input devices cap the polled ports. Automatic (the default) keeps the historical one-joypad packet and grows to the highest configured peripheral port; 0 reports an empty chain.
  • Madam now uses Opera's 256-byte PBUS buffer. The state section is version 4; v1-v3 states read the historical 20-byte queue and fill the tail with 0xFF.
  • Tests: EightPlayerPacket_ReachesTheMadamPbusQueue sends eight joypads through FiveDoMachine and checks the eighth record at queue offset 14; EightJoypads_UseTheOperaDaisyChain checks the 20-byte chain packet; ApplyInputDevices_ActiveCountControlsPolledPorts covers 0, 3, and 8; SaveState_RoundTripsTheExpandedPbusQueue covers the 256-byte queue; KeyPressed_PreservesAnEightDeviceChain checks the tail fill.
  • Hardware/test-ROM six-player and multi-tap verification remains; this is model-level packet and machine-path evidence, not a hardware capture.

1.1an SYS-12 Portfolio OS SWI HLE (2026-11-04)

  • ArmSwiHle ports Opera's 14 intercepted SWIs. 0x50003, 0x50004, 0x5000A, 0x5000B, and 0x5000D return false so the normal guest trap runs.
  • Tests: ArmSwiHleTests covers every intercepted function, the matrix/vector layouts, perspective divide, the sqrt_frac16 restoring loop, unsupported-SWI fallback, the disabled trap, equal cycle cost for the SWI itself, and telemetry reset. The expected values for MulVec3Mat33, MulMat33Mat33, MulVec3Mat33DivZ, MulVec4Mat44, MulManyVec4Mat44, and MulMat44Mat44 were generated with References/opera-libretro-master/libopera/opera_fixedpoint_math.c and match bit-for-bit.
  • SwiHle_IsWiredToMachineRam executes SWI #0x50000 through FiveDoMachine and checks the result at the guest destination.
  • Options: swiHle is registered in MachineOptionRegistry, round-trips through Settings, per-game profiles, and save-state Options v9, and appears in docs/MachineOptions.md.
  • CLI: --swi-hle enables the mode; --swi-comments emits a per-comment histogram.
  • Real-BIOS benchmark (4,096 MulMat44Mat44_F16 calls): 847,860 LLE steps / 4,054,999 cycles / 0.025 s versus 24,578 HLE steps / 65,538 cycles / 0.002 s. See Performance.md §8.2.
  • Title evidence: AutoBahn Tokio (3,891 calls in 3,000 frames) and Ballz (2,880 calls) boot and run with HLE. Their frame hashes differ from the default because HLE skips the guest handler's cycles/register clobbers; leaving the affected SWI to the guest restores the hash. Lemmings has no HLE calls and the differential smoke is an exact match. Default-off behavior is unchanged.

1.1ao SYS-17 game database expansion (2026-11-05)

  • GameDatabaseNameParser parses Redump, GoodTools, and No-Intro filenames, including release dates, compound … & … titles, region lists, language lists, disc numbers (Disc 2 of 3, Disk B), revisions (Rev 1, RE2, v2.2), and dump/edition tags. Unknown parenthetical groups stay in the title; uncertain fields stay empty.
  • GameDatabase now parses and writes revision, discNumber, discTotal, bios, bios2, fixMode, inputDevices, nvram, and cheats. Save/Load roundtrip the full record, and Merge preserves curated records unless --overwrite is supplied.
  • gamedb stats|scan|export|merge are covered by CLI tests. scan computes IDs for .chd/.cue/.iso files and writes a fragment; merge adds it to the embedded database and writes the reviewed result. Transient CHDSharp open failures are retried serially.
  • Runtime integration: GameCompatibility.ForGameId reads the database fixMode; BiosLibrary.SelectForGame reads bios/bios2 (per-game override still wins); and GameControlDefaults reads inputDevices. info --json exposes the new fields.
  • Local coverage: gamedb stats --folder "J:\Panasonic 3DO" reports 1,366 images found, 1,366 readable, 653 unique IDs, 1,366 matched, 100% coverage, and no errors. The three former exceptions were audio-only [CD-Audio] CHDs, which now open with TOC-derived IDs.
  • Embedded database: 518 original 4DO records + 135 local records = 653. Known BIOS requirements (Powers Kingdom Europe, WorldCup Super Stadium Japan + Kanji), fix masks (Crash 'n Burn, Street Fighter II, DinoPark Tycoon, Microcosm), and lightgun/Orbatak input defaults are stored as data.

1.1ap SYS-08 per-title CPU clock behavior (2026-11-05)

  • Five titles were run for 3,000 frames at 0.5x, 1x, 2x, and 4x with --audio null. Every non-stock clock produced a different frame hash, so the clock is a compatibility variable rather than a speed switch. AutoBahn Tokio took 9.1 s at 1x, 84.4 s at 2x, and 271.7 s at 4x for the same 3,000 frames.
  • The stock 12.5 MHz default, the synthetic clock-invariant benchmark, and the decision not to seed per-title clock overrides are recorded in Performance.md §8.3.

1.1aq SYS-01 audio-only CHD increment (2026-11-05)

  • ChdDiscImage now opens a CHD whose track table has no data track. ReadTrackSector returns the native 2352-byte audio sectors; ReadSector still rejects data reads from an audio track.
  • GameId.Compute gives an audio-only disc a stable MD5 derived from its track table of contents (audio-cd|track:sectors|...) instead of failing on the missing data sectors. The same TOC yields the same ID; a changed track length yields a different ID.
  • info reports audioCd = true and trackCount and skips the volume-header read. The three local [CD-Audio] discs (A&M Music Sampler, D's Diner - The Director's Cut, Oyaji Hunter Mahjong) are now catalogued in the game database.
  • The BIOS still requires a data track to boot, so the audio-only discs are not yet playable in-emulator; PLAY_AUDIO_TRACK and the CD-DA mix remain covered by the synthetic CdromMultiTrackTests/CdAudioMixTests.

1.1ar SYS-18 auto per-game profiles (2026-11-05)

  • AppSettings.ResolveInputDevices resolves a per-game override, then the game-database inputDevices spec, then the global list. CLI run/trace apply the database spec when --input-device is absent. EnsureAutoInputDeviceOverride skips database titles so a generated override cannot shadow future database updates.
  • SettingsWindow changes flow through one CreateEffectiveForGame snapshot. Input, filter, and high-resolution changes apply live; BIOS, NVRAM, fix, and timing changes recreate the machine. GameProfile.Clone copies the device list, and the profile carries a CheatBundle name.
  • Tests: database input resolution for Crime Patrol (E8FBE61C) and Orbatak (B6B3F5B2), no shadow override, profile clone independence, and immediate effective-profile changes for filter/input/fix/BIOS.
  • Manual evidence, all with --bios-auto bios and no manual BIOS/region/fix/input flags:
    • WorldCup Super Stadium (Japan): selected Panasonic FZ-10J + Kanji ROM (c23fb5d5e6bb1c240d02cf968972be37) and rendered the "Story of the World Cup" screen at 3,000 frames.
    • Powers Kingdom (Europe): selected Panasonic FZ-10E (Anvil), region=Pal1, fieldRate=50, sourceHeight=288.
    • Slayer (USA): selected Panasonic FZ-10.
    • Crash 'n Burn (USA): applied fixMode=1 (FIX_BIT_TIMING_1).
    • Crime Patrol (USA): selected 1=lightgun from the database and rendered its live-action intro at 3,000 frames.

1.1as PERF-04 prefetch buffer increment (2026-11-05)

  • PrefetchingDiscSource keeps a bounded pool of evicted sector buffers, so a full cache reuses its arrays instead of allocating a new byte[] per stored sector. The worker also reuses one 2048-byte data buffer and one 2352-byte audio buffer.
  • StoreData/StoreTrack copy into pooled buffers; eviction returns the old array to the pool; Dispose clears the cache and pool.
  • CacheHits_DoNotAllocateOnTheCallerThread warms a cached sector and then performs 1,000 cache-hit reads, asserting GC.GetAllocatedBytesForCurrentThread() stays below 128 bytes (a one-time runtime allocation is tolerated).
  • This is an increment, not completion: continuous miss/eviction streaming still needs an end-to-end allocation gate and frame-spike measurement.

1.1at PERF-08 managed IL JIT (2026-11-05, partial)

  • Arm60.Jit.cs compiles hot straight-line ARM60 blocks into DynamicMethod IL. Only data-processing, immediate-offset load/store, and branch instructions are compiled; multiply, swap, block transfer, SWI, MRS/MSR, mode changes, PC writes, and register-specified shifts stay on the interpreter.
  • Stores and branches terminate a block. Every compiled instruction is verified against the current code word before it executes; a changed first word returns a stale sentinel and a changed later word returns the committed cycles so the next call recompiles or falls back. This covers self-modifying code, DMA, ROM-bank switches, and state loads.
  • Blocks stop at the frame loop's 16-cycle batch boundary using the machine's current batch, so CLIO/CD/DSP/VDLP event timing matches the interpreter. Blocks compile only after 8,192 executions to keep one-shot boot/decompression code on the interpreter.
  • ArmJitTests covers data processing, conditions, loads/stores, branches, link branches, LSL carry-out, stale-code refresh, synthetic boot, and a machine-level frame/RAM parity test. The CLI test compares interpreter/JIT frameHash and instruction counts.
  • Lockstep evidence: 100,000,000 Lemmings instructions with identical PC, CPSR, registers, cycles, and instruction count. Full-frame hashes match for Lemmings 3,000 frames (0x9C2D31F6), Road Rash 1,000 frames (0x41D2AC9A), and Powers Kingdom Europe 1,000 frames (0xDCEB4C05).
  • Performance is a partial result: ALU/load microbenchmarks are 1.9-3.3x faster, but end-to-end boot is 1.17-1.85x slower because short blocks, CEL/VDLP work, and JIT code-cache pressure dominate. The interpreter remains the default. See Performance.md §8.4.

1.1au SYS-06 RSA/unsigned-disc option (2026-11-06)

  • BiosRsaPatch implements the community documented two-site media-RSA bypass: each 000313A0000111A0 sequence becomes 000313A0000011A0; the 000213A0000111A0 guard sequence is left unchanged.
  • Local verification with the retail images:
    • panafz10.bin (51f2f43ae2f3508a14d9f56597e2d3ce) + patch -> 1477bda80dc33731a65468c1f5bcbee9, exactly panafz10-patched.bin.
    • panafz10e-anvil.bin (a48e6746bd7edec0f40cff078f0bb19f) + patch -> cf11bbb5a16d7af9875cca9de9a15e09, exactly panafz10e-anvil-patched.bin.
  • run --bios bios/panafz10.bin --allow-unsigned-discs --frames 0 --json reports biosMd5: 1477bda80dc33731a65468c1f5bcbee9, biosRsaPatched: true, and biosRsaDisabled: true; the source BIOS file is unchanged.
  • A 600-frame Wolfenstein 3D (USA) run produced the same frame hash 0x16B5C318 (381,025,496) with the retail BIOS and with the runtime-patched BIOS; a 600-frame Road Rash (USA) run produced the same hash 0x6CDB8B35 (1,826,602,437) with the retail FZ-1 and the runtime-patched FZ-1, so signed software is unaffected.
  • bios select "Wolfenstein 3D (USA).chd" --folder bios --allow-unsigned-discs selects Panasonic FZ-10 (No RSA) (1477bda8...); the same command without the flag selects retail Panasonic FZ-10 (51f2f43a...).
  • Unit tests cover patch counting/idempotence, the guard sequence, in-memory-only patching, the no-site error, known-No-RSA catalog flags, a No-RSA file-name fallback, BIOS selection preference, settings round-trip, CLI JSON reporting, and bios info patch-site reporting. bios/ local-file tests skip when the images are absent.
  • The feature is complete: the BIOS patch is explicit and opt-in, the source image is never modified, known No-RSA images are preferred, and a compatible retail image is patched only when requested. No BIOS or game content is bundled.

1.1av VID-12 latency options and PERF-02 AOT (2026-11-06)

  • bench --frames 120 --latency (Release, this desktop): snapshot 9,348,438 bytes, capture 3.6 ms/frame, restore 2.0 ms/frame. The synthetic bench hash remained 0x6CDFBDC5. The uncompressed state is smaller than the previous ~10 MB because Machine section v2 writes only visible lines; RAM restore no longer clears the unused 22 MB tail.
  • LatencyTests (10 tests) covers snapshot frame/hash round-trip, 1- and 2-frame run-ahead timeline parity against a normal run (machineFrames = displayed + depth), visible-only audio, frame-delay ring holding, range validation, and telemetry mean/max.
  • Frontend Debug and AOT win-x64 runs of Lemmings (USA) with run-ahead 1, frame delay 1, VRR, and audio sync 40 ms logged the applied options, opened the disc, started emulation, and produced no core fault. The AOT run also logged OpenAL audio ready (44100 Hz stereo, 100 ms, sync 40 ms, device default) with ALSOFT_DRIVERS=null.
  • AOT CLI: --version 69 ms, bench --frames 2 65 ms, 3,000-frame Lemmings boot ok: true, PC 0x0012B36C, frame hash 0x0C5B2F8E, warnings: []; 240-frame realtime OpenAL run with --audio-sync 40 reported 0 dropped samples, 0 underruns, and frame hash 0x6CDFBDC5 matching the IL build. AOT frontend: 34.5 MB single file, --version 70 ms, smoke 1.2 s; trimmed frontend: 27.9 MB single file, smoke 3.2 s; trimmed CLI: 19.4 MB. See Trimming.md and Performance.md §8.7.

1.1aw TEST-07 reference oracle and five-title panel (2026-11-06)

  • benchmarks/diff/samples5.json runs the local D:/Samples/3DO corpus: Alone in the Dark, Lemmings, Road Rash, Space Ace, and Wolfenstein 3D. Settings: 300 frames, checkpoints 60/120/300, fixed seed 1488, accuracy profile, cdSpeed=unbounded, noGameFixes=true, PC/CPSR/CD comparison, sample tolerance 512, no images. The debug core is artifacts/diff/opera-build-current/OperaCore.dll, built from artifacts/diff/opera-build with MinGW gcc/g++ 15.2.0.
  • Determinism: two consecutive runs with --force produced identical checkpoint arrays for both the 5DO and Opera captures (Compared with ConvertFrom-Json over the capture JSON). This is the fixed-seed gate that the previous wall-clock std::rand/prng16/prng32 seeding could not provide.
  • Verdict: 0 match, 5 diverged, 0 error. The report records firstDivergence; the console, JSON, and Markdown all show it.
  • First divergence: ARM PC at frame 60 on all five titles: Alone 0x000057f8 vs 0x00005954; Lemmings 0x00005aa0 vs 0x00005960; Road Rash 0x000057e0 vs 0x0000597c; Space Ace 0x0000595c vs 0x00005948; Wolfenstein 0x0000594c vs 0x0000596c. Maximum absolute PC deltas across the run: Alone 21,020; Road Rash 7,644; Space Ace 6,084; Wolfenstein 3,172; Lemmings 320 bytes.
  • Lemmings fields:
    • f=60: frame hash matches 0x110e7dc5; PC 0x00005aa0/0x00005960; CPSR 0x600000d3/0x800000d3 (FIQ-disable bit); CD commands 9/9, hash matches; samples 44,101/44,144 (within the 512 tolerance).
    • f=120: PC 0x00005988/0x00005974; CPSR matches 0x800000d3; CD commands 50/50, hash matches; samples 88,202/88,288.
    • f=300: PC 0x000205f4/0x00020600; CPSR matches 0x00000010; CD commands 108/105, hash diverges; samples 220,508/220,720.
  • CD capture semantics: 5DO counts the opcode when the first command byte is written (CommandOpcodeObserver); the reference counts XBP_SET_COMMAND at the same phase. The earlier 7x count difference was a measurement artifact (the reference hook fires per command byte); counting only the first byte of each seven-byte packet removed it. A command in flight at a checkpoint can still leave a 1-3 count difference.
  • Audio: the cumulative audio hash is recorded but not gated. The 59.94 Hz reference field rate and 60 Hz 5DO field rate produce a systematic sample-count offset (43 samples at f=60, 212 at f=300), so the streams must be aligned or resampled before the hash is comparable.
  • 3,000-frame follow-up (samples5-long.json): same five titles, 3,000 frames, checkpoints 60/300/1000/3000, no images. Lemmings matched the frame hash at all four checkpoints despite PC drift. Wolfenstein and Alone had transient frame-hash mismatches (Wolfenstein f=300; Alone f=1000) and recovered by f=3000. Road Rash and Space Ace stayed visibly divergent: f=1000 and f=3000 images show different animation frames/scenes, not a rendering error. A ±60-frame offset probe on Road Rash found no neighboring hash match, so the divergence is larger than one second of fields or the frame selection differs. CD command counts stayed close but not exact (Road Rash 611/598 at f=1000; 851/836 at f=3000). This is the first evidence that separates "PC drift with matching video" from "visible divergence" and gives CORE-04 and CORE-13 a concrete lead.
  • Full CD command sequence (cdCommands): the final-checkpoint opcode sequence localizes the first divergence beyond count drift:
    • Lemmings command 105: 5DO 0x10 (READ_DATA) vs Opera 0x83 (READ_ID); 5DO 364 opcodes, Opera 352.
    • Road Rash command 559: same pattern; 851/836.
    • Wolfenstein command 389: same pattern; 862/861.
    • Alone command 433: equal counts (932/932) but a shifted 0x10/0x83.
    • Space Ace: prefixes match for 348 commands; 5DO has one extra 0x10. This localizes CORE-04 to READ_DATA completion/read-ahead: after a few READ_DATA commands, Opera returns to READ_ID polling while 5DO keeps issuing READ_DATA. The earlier transcript-only build used a 7x byte count; the current export counts one opcode per seven-byte packet.
  • Panel generator: panel --from benchmarks/diff/full.json --out benchmarks/diff/panel50.json --count 50 generated a 50-title panel from the 594-disc manifest. It selects evenly spaced titles (first and last included), preserves roots/BIOS/machine options, clears per-title frame overrides and allowlists, and applies the TEST-07 comparison policy. validate accepts it. The full panel has not been run because the J:/G: corpus is not mounted on every machine.
  • Tests: 1,881 pass. New coverage: HarnessCommonTests (default seed, manifest hex seed, override precedence, random rejection), DiffComparerTests (CPSR gate, CD count/hash gate, full CD sequence divergence/prefix, sample tolerance, earliest first divergence), CdromTranscriptTests (issue-time hash/count, opcode list, and reset), and DiffManifestBuilderTests (panel selection, policy, and count edge cases).
  • Limitations: five titles, not the 50-title panel; no scripted input yet; the audio hash is informational. See DifferentialHarness.md §8/§9 and ImplementationPlan.md TEST-07.

1.1ax TEST-07 exact clock and 50-title panel (2026-11-06)

This entry supersedes the clock and first-divergence claims in 1.1aw. Both statements were measurement artifacts of 5DO's 60.00 Hz NTSC clock:

  • Corrected CPSR reading: 0x600000d3 vs 0x800000d3 differ in the N/Z/C condition flags (bits 31-29). The FIQ-disable bit (bit 6) is set in both values. The frame-60 PC/CPSR difference is a frame-boundary phase sample, not evidence of an FIQ-ordering bug.
  • Exact NTSC field rate. NTSC now runs at 60000/1001 (208,541 cycles per field at 12.5 MHz) instead of 12,500,000/60 = 208,333. The DSP sample clock therefore produces 44,145 samples at frame 60 vs Opera's 44,144 (was 44,101), 220,728 vs 220,720 at frame 300 (was 220,508/220,720), and 2,207,289 vs 2,207,207 at frame 3,000. Every one of the 50 panel titles is within 9 samples at frame 300.
  • Rational VDL accumulator. The VDL clock no longer adds the per-batch rounded 8.24 ARM fraction, which ran the scanline clock ~0.09% fast. Quartz accumulates cycles * fieldLines * fieldRateNumerator against armClock * fieldRateDenominator; 60 NTSC fields produce exactly 15,780 scanlines. Quartz state version 4 stores the exact accumulator; v1-3 states convert their 8.24 fraction on load.
  • 50-title panel (artifacts/diff/test21-panel50/report.json). 50/50 captured, 0 errors. 32/50 titles match every frame hash at frames 60/120/300; 17 match one of three, and one bonus disc (Icebreaker II) has no video. 49/50 match the CD command count at frame 60 (delta -1..0); at frame 120 14/50 match (delta -2..0) and at frame 300 24/50 match (delta -2..+4). The first divergence is PC at frame 60 on 48/50 titles.
  • Instruction parity (new). The reference core exports sg_debug_arm_instructions; the harness records it beside 5DO's instructions. At frame 60 Opera's absolute count is 89,338-90,063 instructions higher on all five sample titles (Lemmings 5,383,387 vs 5,472,729), while the sample counts match. The offset is already present at frame 1 and stays fixed; the per-frame retirement rate matches within 0.1% over 240 frames on 4/5 titles, with Lemmings diverging only where its CD path diverges. This is a startup/first-field capture offset, not a cycle-model error, so it does not implicate CORE-01. FreeDO N=4 ARM timing is still far worse (3,855,498 instructions).
  • CD cycle stamps (new). The debug core exports sg_debug_cd_cycle_at, recorded at the exact command boundary (cmd_idx == 0, also fixing the old byte-modulo-seven transcript for one-byte ABORT packets). The first 105 Lemmings opcodes match, but 5DO reaches command 104 at cycle 57,736,851 vs Opera's 57,621,696 (+115,155 cycles, ~9.2 ms). Road Rash +16,680, Space Ace +80,866, Wolfenstein +13,343, Alone +15,163 at their last common command. The remaining CORE-04 lead is a cadence/READ_DATA-completion difference, not the command table.
  • 3,000-frame recheck. With the exact clock, Space Ace matches Opera's frame hash at 3,000 (0x46916c61). Wolfenstein (0x175b8fe4), Alone (0x175b8fe4), and Lemmings (0xebdb01ca) also match at 3,000. Road Rash still diverges at 1,000 (0x63165fe3 vs 0xfefd249a) and 3,000 (0xd5895ad5 vs 0x0528251f). The blanking/background fix did not change these long-run keys.
  • Tests: 1,885 pass. New coverage: exact-rational frame cycles, 60 fields = 15,780 scanlines, sample count 44,144 +/- one, state v4 round-trip, behavioral divergence classification, and CD cycle capture. dotnet format is clean.
  • VDLP blanking and background color (new). Opera renders a black line when bitmap DMA is disabled; 5DO kept the previous field's pixels. Vdlp.DoLine now clears the visible pixels and background in both resolution paths. The background command parser also byte-swapped r/b and mixed the green channel; it now narrows the command's 8-bit r/g/b to the 5-bit line format used by the assembler and GPU. The green boot field became black, matching the reference's near-black; the remaining hash difference is Opera's 8-bit background precision (1,1,1) versus 5DO's 5-bit (0,0,0), not a game-visible defect.
  • Final long-run harness report (artifacts/diff/test23-long/report.json). Same five titles, 3,000 frames, checkpoints 60/300/1000/3000, final build. Lemmings matches at all four checkpoints (0x110e7dc5, 0x110e7dc5, 0x110e7dc5, 0xebdb01ca); Alone and Wolfenstein match at 60/300/3000 (Wolfenstein's 1,000 also matches); Space Ace matches at 60 and 3,000 (0x46916c61); Road Rash matches 60/300 but still diverges at 1,000 (0x63165fe3 vs 0xfefd249a) and 3,000 (0xd5895ad5 vs 0x0528251f). Sample deltas stay under 100 at 3,000 frames (52-86), and the CD counts are within 1-13.

1.1ay 50-title video-clean panel, instruction-phase triage, and VDL fixes (2026-11-07)

This entry supersedes the video and PC-phase claims in 1.1ax. The 50-title panel now matches the reference on every frame hash, and every PC/CPSR difference is measured as a frame-boundary instruction phase.

  • Panel command. FiveDo.DiffHarness.exe run artifacts/diff/panel50-local.json --out artifacts/diff/test43-panel50-bg --no-images --compare-cd --core artifacts/diff/opera-build-local/OperaCore.dll, with the Release CLI built from the working tree, fixed seed 1488, accuracy profile, cdSpeed=unbounded, and noGameFixes=true.

  • Result. 50/50 captured, 0 errors. 50/50 titles match every frame hash at frames 60/120/300 (up from 32/50). 50/50 first PC divergences instruction-align with Opera frame 59; the residual is only -55 to -256 instructions, while the same-frame delta is about 89,000. The only remaining behavioral divergence is cdCommandCount: 36 titles first differ at f=120 (1-2 commands in flight; all recover by f=300), 13 at f=300 (seven FMV titles at +3/+4 extra READ_DATA commands, six at -1/-2), and Icebreaker II at f=60.

  • Reason for the video change. Two VDLP bugs were masking the reference match. The background register is 24-bit RGB; 5DO narrowed it to 5 bits and then expanded it, turning the reference's near-black (1,1,1) into (0,0,0). The CPU assembler, GpuPalette, the OpenGL shader, and the Direct3D shader now all keep the full eight bits per channel. Separately, a zero VDL display-list word is a stop that retains the previous control word and bitmap (opera_vdlp.c); 5DO treated it as an entry and blanked the last visible scanline. The machine VDL frame section is version 3 and converts version 1-2 5-bit backgrounds on load.

  • Five-title panel. test39-vdlstop-samples5 matches 5/5 frame hashes at 60/120/300, including Space Ace (0x76175dc5 at f=120, 0xde7d0ba6 at f=300) and Wolfenstein 3D (0x76175dc5 at f=120 and f=300).

  • 3,000-frame panel. test44-long-bg: Alone, Lemmings, Space Ace, and Wolfenstein match the frame hash at 3,000 (0x175b8fe4, 0xebdb01ca, 0x46916c61, 0x175b8fe4); Road Rash still diverges at 1,000 (0x63165fe3 vs 0xfefd249a) and 3,000 (0xd5895ad5 vs 0x0528251f). CD counts are Lemmings 365/352, Road Rash 851/836, Space Ace 349/348, Wolfenstein 862/861, and Alone 932/932. Alone and Space Ace have transient video mismatches at f=1000 that recover by 3,000.

  • Harness diagnostics. Each report now records pcPhaseFrame and pcPhaseInstructionDelta; the Markdown/HTML tables show a Phase column, and the detail text names the aligned reference frame. ISO staging is atomic (.tmp + move), retries once after a transient CHD decompression failure, and a per-title capture exception becomes a title Error instead of aborting the panel. A /tmp-style partial ISO can no longer be reused by --resume.

  • Johnny Bazookatone check. The zero-entry VDL stop does not change Johnny Bazookatone (USA) at 3,000 frames; the pre- and post-change 5DO hashes are both 0xa29ad2d1 (the title's reference hash remains 0xcf5644e7, a pre-existing CORE-13 divergence).

  • Tests: 1,916 pass. New coverage: instruction-phase neighbor matching and summary counts, 24-bit background expansion in both palette paths, eight-bit background preservation, VDL frame state version 3 round-trip, and atomic ISO extraction. dotnet format is clean and InspectCode reports 0 results.

  • CORE-04 READ_DATA validation (2026-11-07). CdromDevice.ReadDataCommand now rejects an invalid MSF with MEI address error, rejects a transfer that starts at or runs past the last disc sector with MEI end address, and aborts a previous transfer before starting a new one. Three tests cover both error paths and the stale-sector abort. The five-title and 20th Century counts are unchanged. A separate oracle build with the current Opera transfer-completion read-status rule also produced the same counts, ruling out the local oracle's first-sector publication and past-end reads as the cause of the residual FMV cadence; the next lead is the one-field boot phase and the guest's FMV read loop.

1.1az READ_DATA first-sector status, deadline accumulation, and per-command PC/instruction stamps (2026-11-07)

This entry supersedes the count-only conclusion at the end of 1.1ay. The new per-command stamps show that the publication rule changes the boot-handoff timing even when the 300-frame command count is unchanged.

  • Oracle/5DO instrumentation. The debug core exports sg_debug_cd_pc_at and sg_debug_cd_instructions_at; CdromCommandTranscript records IssuedProgramCounters/IssuedInstructions at the same command boundary as the cycle stamp. --cd-commands now writes opcode cycle pc instructions, and captures carry cdCommandPcs/cdCommandInstructions.
  • 20th Century probe. artifacts/diff/test51-20th-boot-probe (before) and test52-20th-first-sector (after), 300 frames, accuracy profile, unbounded CD speed, artifacts/diff/opera-build-pc/OperaCore.dll. At command 50 (the last boot READ_DATA to READ_ID handoff) 5DO spent +207,707 cycles / +86,511 retired instructions with the same guest PC (0x00022140) as the reference. Publishing POLST with the first queued READ_DATA sector, matching cdrom_data_publish_read_status_if_ready (data_len != 0), reduced the deltas to -818 cycles / -376 instructions. The f=300 command-104 cycle delta fell from +324,027 to +111,990.
  • Deadline accumulation. Each later READ_DATA sector now adds one sector-time to the previous deadline (cdrom_data_schedule_next) instead of re-basing to now + sector_cycles. OperaCycleBased_AccumulatesSectorDeadlinesWhenTheGuestDrainsLate covers a half-sector late drain.
  • No video regression. test53-samples5-first-sector keeps 5/5 frame hashes at 60/120/300 on the five-title sample; test57-samples5-x2-final keeps 5/5 at the stock 2x drive speed. test56-long-first-sector keeps the established 3,000-frame result: Alone, Lemmings, Space Ace, and Wolfenstein match the final frame hash, Road Rash still diverges, and the CD counts are Alone 932/932, Lemmings 365/352, Road Rash 851/836, Space Ace 349/348, and Wolfenstein 862/861. 20th Century (test58-20th-3k) matches all four video hashes through 3,000 frames and ends at CD 332/311. The 50-title count gate is unchanged at 300 frames; the improvement is in cycle/instruction phase, not in the checkpoint command count.
  • 50-title confirmation panel. test55-panel50-first-sector captured 50/50 titles with 0 errors and 50/50 match every frame hash at 60/120/300 using artifacts/diff/opera-build-pc/OperaCore.dll (MD5 53c26d4cb701e3d6e3bb28e877cd502b). 48/50 first PC divergences align with Opera frame 59 and 2/50 with frame 119 (those two matched at f=60). The first behavioral divergence is still only cdCommandCount: 36 at f=120, 13 at f=300, 1 at f=60 (Icebreaker II), identical to test43. One title hit a transient CHD read error in the first pass; --resume retried it and the final report is error-free.
  • Tests: 1,919 pass in Debug and Release. New coverage: first-sector POLST publication, accumulated deadlines, and PC/instruction capture in the CD transcript. dotnet format is clean and InspectCode reports 0 results.

1.1ba Frontend fullscreen launch argument (2026-11-07)

-f, -fullscreen, and --fullscreen are equivalent. Program.IsFullscreenRequested has a unit test, and --smoke-test --fullscreen now applies the launch mode before writing the marker:

FRONTEND SMOKE OK 1.4.0
windowState=FullScreen

The normal smoke run records windowState=Normal; the marker's first line is unchanged for existing CI consumers. 1,919 tests pass.

1.1bb CD transcript classification and cadence-step triage (2026-11-07)

The differential harness now classifies the final CD transcript and reports the largest cadence step in the cumulative 5DO-minus-Opera cycle delta. cdCadence carries the classification (Equal, PrefixAhead, SequenceDiverged), the first mismatch, the last matched cycle/instruction delta, and the step's opcodes, cycle jump, command-issue PCs, and retired-instruction jump. The Markdown and HTML reports add a CD cadence column and detail line.

  • 50-title panel (test55-panel50-first-sector/report.json). 39 identical streams, 8 prefix-only, 3 sequence-shifted. The three shifts are one READ_ID poll among READ_DATA commands: Chiki Chiki 111/112 (extra Opera READ_ID), Dragon's Lair 111/108 and Virtual Horror 110/107 (extra 5DO READ_ID). 24/50 titles have a step above 1,000 cycles; every one has the same command-issue PC on both engines (17 at 60-100k, six at 20-60k, one >=100k). The 60-100k group is a single READ_DATA-to-READ_DATA step where 5DO retires about 32k more instructions in the same guest path.
  • Five-title run (test60-cadence-samples5/report.json). Alone 361/361 (+379 cycles at the last command), Road Rash 395/395 (+468), Wolfenstein 341/341 (-345), Lemmings 109/105 with one +22,834 step at command 51 (PC 0x00022140 both sides, +8,242 instructions), and Space Ace 168/168 with one +80,567 step at command 164 (PC 0x0002aa18 both sides, +32,330 instructions).
  • Register trace (--cd-register-trace, 2026-11-07). The new trace records every CD command write, poll write/read, status read, data write, and empty data read with cycle and PC. Space Ace drains the READ_DATA status by cycle 58,934,229, then has no CD register access from 58,935,434 to 59,173,228 (237,794 cycles); command 164 follows at 59,174,910. Lemmings drains all 12 READ_ID status bytes by 29,431,952, then has no CD access until 31,599,968 (2.17M cycles). Both large cadence steps are guest waits on a non-CD source (timer/VBlank), not CD readiness. This moves the residual cadence lead from CORE-04 to CORE-07.
  • Interpretation. The command table matches on 47/50 panel titles; the residual cdCommandCount difference is guest wait-loop cadence, not a divergent CD command/status model. The next measurement is a CLIO timer/VBlank trace around the step. benchmarks/diff/samples5.json now resolves the local corpus at J:/Panasonic 3DO.
  • Tests: 1,922 pass. New coverage: cadence-step detection with matching PC and instruction jump, sub-threshold step suppression, transcript-kind assertions for equal/prefix/diverged streams, and CD register-trace output for command/poll/status accesses. dotnet format is clean.

1.1bc Frame-boundary parity and checkpoint capture order (2026-11-07)

Two independent defects produced the universal frame-60 PC/CPSR phase:

  • Capture order. OperaCoreHost stored the checkpoint frame number in the video callback but finalized the sample at the next callback. That read the ARM PC/CPSR/instruction count and the CD status after the following field had executed, while _pcm already held the audio through the checkpoint field. The harness now captures ARM/CD state in the checkpoint's own video callback and fills the audio hash when that field's audio callback completes.
  • Field boundary. FiveDoMachine.RunFrame stopped when the rounded ARM-cycle budget (208,541 cycles) elapsed. Opera stops when the scanline counter reaches opera_region_scanlines() (263 lines for NTSC), carrying the remaining cycles in g_FRAME_CYCLE_REMAINDER. The last 16-cycle batch overshoot was discarded, so 5DO accumulated about 3.3 extra retired instructions per field. RunFrame now ends on a change of Quartz.VdlFieldIndex, with a two-field safety cap.

Five-title sample (test63-line-boundary/report.json, debug core opera-build-pc, MD5 53c26d4cb701e3d6e3bb28e877cd502b):

Frame Retired-instruction delta (Opera − 5DO) Notes
60 (5 titles) 8-17 was ~89,000 before the capture fix, ~219 after it
120 (5 titles) 24-84 PCs match within one instruction
60 (50 titles) 11-20 average 15.3; 50/50 frame hashes
120 (50 titles) 23-2,223 most titles around 200; 50/50 CD count match
300 (50 titles) CD cadence only 11 titles: 8 count, 3 one-command sequence

The 50-title confirmation panel (test64-panel50-line-boundary/report.json) captures 50/50 with 0 errors and 50/50 frame hashes. The CD count matches on 50/50 titles at f=60 and f=120, down from 36 first divergences at f=120; the first behavioral divergence moved to 11 titles at f=300 (8 cdCommandCount differences and 3 one-command cdCommands sequence differences). The other 39/50 titles are phase-only at the checkpoints. The 3,000-frame showcase keeps 4/5 video matches (Road Rash is the only frame-hash divergence) with CD counts Lemmings 365/352, Gex 548/547, Road Rash 851/836, Wolfenstein 862/861, and Alone 932/932 (Alone's sequence hash differs but the video matches).

The residual is below one 16-cycle batch. This supersedes the earlier interpretation that 5DO frame 60 aligns with Opera frame 59; that one-field offset was a harness sampling-order bug. DiffComparer now labels a same-frame delta of at most 256 retired instructions as same-frame phase (pcPhaseFrame = the checkpoint frame), so future reports carry pcPhaseTitles for the small residual instead of leaving it untriaged.

The first remaining behavioral divergence on Lemmings is still the CD cadence step between commands 50 and 51: the READ_ID-to-READ_ERROR wait is 2,169,618 cycles in 5DO and 2,146,784 in Opera (+22,834). --fiq-trace has no FIQ events and --cd-register-trace has no CD accesses during the wait. --clio-timer-trace shows timer 14 underflowing 209 times at roughly 10,400 cycles per period; the Opera wait is consistent with a roughly 1% shorter period. --clio-timers opera changes the period to about 12,400 cycles and shifts the boot timeline, so the opt-in Opera scheduler is not the fix. The next measurement is the reference timer-14 period/slack (CORE-02).

Tests: 1,923 pass. New coverage: RunFrame_EndsOnTheVdlFieldBoundary runs 60 frames and asserts the VDL field index changes exactly once per frame and the executed cycle count stays within 64 cycles of the exact budget. dotnet format is clean.

1.1az CLIO timer/read trace and CD FIQ parity (2026-11-08)

The CORE-02 follow-up added reference-side CLIO timer, read, XBUS-select/poll, and FIQ-write instrumentation (sg_debug_timer_*, mirrored by the harness --timer-trace) and a 5DO --clio-read-trace. Both engines now write clio-timer and clio-read lines with frame, ARM cycle, and guest PC stamps.

The timer-14 period is not the residual. On Lemmings the accuracy profile selects the Opera scheduler in both engines, and the timer-14 underflows are 12,384 cycles apart in Opera and 12,396 in 5DO (the small difference is the 16-cycle guest batch), with a constant ~285-cycle phase offset. The READ_ID to READ_ERROR wait is still 2,169,618 cycles in 5DO and 2,146,784 in Opera (+22,834), but the instruction delta is +8,694: the guest runs a data-dependent copy/decrypt loop between the commands, not a timer or CD-readiness wait. The first 0x118 (timer 3 counter) read after READ_ID is 29,491,640 in 5DO and 29,469,632 in Opera, and the last reads before READ_ERROR are identical (0x118 = 0x368 at PC 0x00014ce0). The residual is therefore guest work, not the CLIO period; it is tracked under CORE-13/CORE-04.

The trace exposed two real reference-parity bugs:

  • CD poll enables. Opera resets the CD poll to CDROM_POLL_INTEN_MASK (0x07) and latches POLRE on the boot disc access (observed 0x87). 5DO reset to 0x80 only, so the guest's read-modify-write (read 0x540, write 0x540) cleared all three enables. After the fix the guest writes 0x87, the CD publishes 0x17 for READ_ID, and the reference and 5DO both show FIQ reason bit 2 at 28.78M cycles.
  • Level-held XBUS FIQ. Opera's opera_xbus_tick asks every device for XBP_FIQ each batch; 5DO only raised FIQ 4 on sector publication. The machine now re-raises FIQ 4 while Xbus.FiqPending in OperaCycleBased mode.

The combination also exposed the old READ_ID media-access heuristic: with the enables restored, its POLMA bit became a permanent level FIQ, and Space Ace's 3,000-frame hash moved off Opera's 0x46916C61. The A/B matrix on Space Ace was: poll fix only match, level FIQ only match, both fixes diverge, both fixes minus the heuristic match. The heuristic is retired; Road Rash's 3,000-frame hash stayed 0xD5895AD5 throughout, and the 5-title sample keeps 5/5 frame hashes.

Evidence: artifacts/diff/test70-timer-reads through test91-panel50-heuristic-retired (local corpus, debug core opera-build-timers/OperaCore.dll). The 50-title panel (test91-panel50-heuristic-retired/report.json) captures 50/50 with 0 errors, 50/50 frame hashes at 60/120/300, 49/50 same-frame PC phase, and 38 phase-only titles. It improves Daedalus Encounter to a full match; the remaining 11 first behavioral divergences are CD-only at f=300 (8 count, 3 one-command sequence). Dragon's Lair remains 111/108. The five-title sample keeps the established 5/5 frame hashes and 4/5 phase-only statuses; Lemmings remains the CD-count outlier (109/105 at f=300) and Space Ace its 168/168 command stream with the +80,567 cadence step. The 3,000-frame sample (test92-samples5-long-final) keeps 4/5 video matches (Alone, Lemmings, Space Ace, Wolfenstein); Road Rash remains the only video divergence. Tests: CdromDeviceTests Opera-poll reset/READ_ID, XbusTests.FiqPending, and ClioTests.ReadTrace cases; 1,925 pass (one unrelated allocation test is flaky under load and passes alone). dotnet format and InspectCode are clean.

1.1ba GPU HQx rule-table passes (2026-11-08)

HQx 2x/3x/4x now runs as a shared GPU rule-table pass in OpenGL/GLES, Direct3D 11, and Vulkan. The compact table is generated from the 4DO hq2x/hq3x/hq4x sources and checked in; the CPU HqxFilter remains the reference oracle. The rule table is validated independently and cross-checked against the CPU path at all three scales, and the synthetic probe frames are pixel-identical on the Direct3D 11 and Vulkan readback paths. On the RTX 4070, 4K HQx presentation measures 0.14-0.34 ms per frame and the GPU pass 0.09-0.17 ms, down from 6.82/10.93 ms CPU-bound. The opt-in WGL and ANGLE GL probe also passes the HQx rule-texture path without fallback.

The same increment fixes the remaining panel-wide video divergence: the 3DO background is kept as 24-bit RGB through the CPU assembler and all GPU palette shaders, and a zero VDL display-list word is a stop that retains the previous bitmap. The 50-title panel remains 50/50 frame hashes at frames 60/120/300 with 0 errors. RunFrame ends on the reference 263/312-line field boundary, the harness samples Opera's ARM/CD state in the checkpoint's own callback, and the CD poll/FIQ/READ_DATA parity fixes are recorded in 1.1az. 1,939 tests pass; build, format, and InspectCode are clean.

1.1bd TEST-07 audio gating, scripted input, and the nightly baseline (2026-11-08)

The cumulative audio hash is now part of the 300-frame differential gate. The harness compares it only when both engines produced the same sample count; a count difference is reported as cumulative hash not comparable: sample delta ... (reference 16.16 field-rate quantization, 1.214 ppm) instead of a waveform difference. A hash difference at an aligned count is still a hard failure. The report adds Audio hashes aligned (all checkpoints), Audio lag (sample count), and Audio content diverged summary rows plus an Audio aligned HTML card.

The alignment probe (benchmarks/diff/audio-probe.json, Lemmings, 3,000 frames) measured:

Frame 5DO samples Opera samples Delta
60 44,144 44,144 0
120 88,288 88,288 0
300 220,720 220,720 0
600 441,441 441,441 0
1,000 735,735 735,735 0
2,000 1,471,470 1,471,471 +1
3,000 2,207,205 2,207,207 +2

5DO follows the exact 60000/1001 rate (59.940059940 Hz). OperaCore's 16.16 constant is 3,928,227/65,536 (59.939987183 Hz), 1.214 ppm lower; a field is 0.253 ARM cycles longer, or 0.000893 samples per field. The predicted drift at 3,000 frames is 2.68 samples, matching the measured +2. This is the reference's clock quantization, not a 5DO sample-clock error, so 5DO does not copy the constant.

The 50-title panel (artifacts/diff/test93-panel50-audio/report.json, local corpus, debug core opera-build-timers/OperaCore.dll) gates audio and captures 50/50 titles with 0 errors, 50/50 frame hashes, and 150/150 audio hashes at frames 60/120/300. The 11 CD-only first behavioral divergences at f=300 remain; all other titles are PC/CPSR phase-only.

Scripted input is wired through both engines. benchmarks/diff/scripts/menu-advance.txt is a generic menu script (Start and A with gaps); samples5-input.json runs the five local titles for 900 frames. Results (artifacts/diff/test94-samples5-input/report.json):

  • 5/5 frame hashes at 60/300/600; every title matches the reference video through the first six seconds of input.
  • The script reaches the guest: 5DO output changes at f=600 for Space Ace and at f=900 for Alone, Lemmings, Road Rash, and Wolfenstein 3D.
  • The no-input control (test96-samples5-noinput) matches the same checkpoints, so the 60/300/600 matches are not from an ignored script.
  • Road Rash's f=600 video divergence reproduces with no input (test95-roadrash-noinput), so it is not input-induced.
  • At f=900 the reference is one sample ahead on all five titles, so the audio hash is reported as the alignment state. Alone's f=900 frames are the same dark transition/noise screen in both engines; the pixel hash differs but the visible output does not.

The nightly workflow (diff-harness.yml) now runs the showcase with --force, uploads report.html/report.md/report.json, and compares the fresh report against benchmarks/diff/baseline/showcase-report.json. A changed 5DO field fails the job; a baseline-known divergence (Road Rash at 3,000 frames) does not. Generate the baseline with FiveDo.DiffHarness baseline --current <report.json> --write <baseline.json>.

Tests: audio content/lag comparison, the report audio summary, and baseline write/compare; 2,043 pass. dotnet format and InspectCode are clean.

1.1be CORE-04 CD poll reset parity and the 900-frame cadence probe (2026-11-08)

Opera's opera_cdrom_set_poll treats POLREMASK (0x08) as a one-shot FIFO reset and stores only the three interrupt enables (0x07) in the poll low nibble. It also does not re-latch POLRE when the reset bit is released. 5DO stored the whole control nibble, so a guest write of 0x08 left the reset bit asserted; SetCommand, GetStatus, GetData, and Tick then dropped or returned zero until the guest wrote poll again. CdromDevice.SetPoll now masks the stored enables, never persists 0x08, and no longer relatches POLRE on release. The defensive reset gates were removed; the load-state migration still applies LocalReset to old states that stored 0x08.

New tests: ResetControlBitDoesNotPersistInPoll, ReleasingResetDoesNotLatchMediaAccess, the updated LocalResetClearsTransferWithoutGatingCommands/MediaAccessLatchIsWriteOneToClear, and Transcript_DoesNotRecordCommandParametersAsOpcodes (a READ_DATA frame byte 0x16 is a parameter, not an opcode). The 5-title 300-frame sample is unchanged: 5/5 frame hashes, Alone 358 commands, Lemmings 109, Road Rash 391, Space Ace 168, Wolfenstein 341, and Lemmings' 300-frame hash is still 0x110E7DC5.

A 20th Century register trace (artifacts/cdreg-20th.txt) shows the guest writes poll 0x87 once and then 0x07/0x35; it never writes 0x08, so this fix is not the 109/105 cadence cause. The new benchmarks/diff/cd-cadence.json probe runs the 11 panel titles that first diverge on CD to 900 frames. The count gap grows: 20th Century 231/211, Creature Shock 242/241, ESPN Fitness 166/154, Dragon's Lair 169/159, Life Stage 143/131, Pebble Beach 179/167, Virtual Horror 166/155, Woody 140/128; Foes of Ali (288/287) and N.O.B. (245/244) are prefix-only, but both also show an aligned-count audio content divergence at f=600. The first opcode mismatch is 5DO 0x10 (READ_DATA) versus Opera 0x83 (READ_ID) on eight titles (Virtual Horror is the reverse), so the residual is a real guest CD read-schedule difference that reaches the audio stream, not a 300-frame in-flight artifact. 2,043 tests pass; build, format, and InspectCode are clean.

1.1bf CORE-04 CD register trace and status/metadata parity (2026-11-08)

The debug reference core now records every CD command byte, poll write/read, status read, and data read with cycle and PC (sg_debug_cd_register_*), and the differential harness writes matched fivedo.cdreg.txt/opera.cdreg.txt files with --cd-register-trace. 5DO records datr for every non-empty data read in addition to datr_empty.

On 20th Century Video Almanac (290 frames, unbounded speed, accuracy profile) the command packets, status bytes, and 2048-byte sector data are identical through command 105, the last command before the sequence divergence; the first mismatch is command 106 (5DO READ_DATA vs Opera READ_ID). The trace exposed four reference divergences now fixed:

  • The XBUS 2x status bit: MEI_CDROM_DOUBLE_SPEED is 0x80, not 2, and cdrom_mode_reset starts the drive at double speed. The drive now reports 0xE3 from reset, the guest's MODE SELECT page 3 byte drives the bit, and the stored mode is re-applied on inject.
  • EffectiveSpeed follows cdrom_effective_speed: unbounded stays unbounded, configured 1x stays 1x, and any configured speed above 1x drops to 1x while the guest has selected single speed. The overhead compensation follows the effective speed.
  • The readable-disc limit is the 3DO volume header's volume_block_count (sector 0, offset 0x50, big-endian), not the physical image size. 20th Century declares 327,168 blocks in a 327,320-sector image; READ_DISC_INFO and the READ_DATA end checks now match the reference.
  • TOC control bytes include the Q-position bit (0x10), and READ_SESSION_INFO returns the reference's 8-byte single-session payload.

The CD-ROM save-state section is version 5 for the guest speed mode and readable limit. The confirmation panel (artifacts/diff/test110-panel50-metadata/report.json, 50 titles, 300 frames, debug core opera-build-timers/OperaCore.dll) captures 50/50 titles with 0 errors, 0 frame-hash divergences, 150/150 aligned audio fields, 39 phase-only titles, and the same 11 CD-only first behavioral divergences at f=300. The remaining 20th Century difference is guest sleep time (6 frames in 5DO vs 60 in Opera) with matching CD values, so the residual is not the CD register contract; the next measurement is the wake source (FIQ/timer/SPORT). The 11-title cd-cadence panel keeps the same first mismatch indices. 2,048 tests pass; build and format are clean. The local InspectCode host crashed before emitting a result in this session; the prior complete run was clean.

1.1bg CORE-04/CORE-07 guest wake-source parity (2026-11-08)

The 5DO --clio-timer-trace now emits the same diagnostic events as the reference debug core: FIQ generation (event=fiqgen), FIQ-register writes (event=fiqwrite), XBUS select/poll (event=xbussel/event=xbuspoll), and XBUS FIQ edges (event=xbusfiq, with the reference source code 1 tick, 2 poll tick, 3 built-in command, 4 device command/poll, 5 built-in poll). The edge rule matches xbus_raise_fiq: the event is written only while CD FIQ bit 4 is not already active.

The wake-source probe (artifacts/diff/probe-20th-wake, 20th Century Video Almanac, 290 frames, accuracy profile, unbounded CD speed) compares the 20th Century wait directly:

  • During the 57.8M-59.0M cycle interval both engines read the active-FIQ register exactly 99 times, with identical distributions: 38x 00000001, 33x 80000001, 23x 80000801, and 5x 80000003. FIQ delivery, masks, and handler cadence are in lockstep.
  • The guest is in the Portfolio OS idle loop at 0x205f4..0x20600 during the wait. The 5DO CPU leaves the loop on the VBlank FIQ at cycle 59,021,271 and issues READ_DATA at 59,063,402. Opera stays idle; its next command is READ_ID at 70,080,576, 12.46M cycles later.
  • On 900-frame captures (artifacts/diff/probe-20th-wake-900) the READ_ID media-poll cadence after the first poll is the same on both engines (Opera 70.08M/82.59M/95.11M/107.62M/120.13M/132.64M; 5DO 70.29M/82.80M/95.31M/107.83M/120.34M/132.85M, a +0.21M cycle phase). 5DO inserts READ_DATA bursts between the polls (59.06M, 59.12M, 61.56M, 64.07M, 66.57M, 69.07M); Opera issues only the polls.

The residual is therefore not the FIQ/timer/SPORT wake source, the CD command table, the CD status bytes, or the READ_ID payload (both return 83 00 10 00 01 00 00 00 00 00 00 e3). It is the guest file-system read-request schedule: 5DO's loader continues to request sectors after command 104 while Opera's loader stops and only runs the periodic media poll. A 2x A/B (artifacts/diff/probe-20th-speed2x, 5DO cdSpeed=2x against the reference's unbounded default) grows the per-command delta by exactly 81,246 cycles per READ_DATA, matching cdrom_data_sector_cycles (104,166 - 2,087); the sector-time formula and the panel's unbounded setting are correct. The next measurement is the FS read request and the CD task's wake queue, not the interrupt controller.

2,051 tests pass (three new trace tests); build and format are clean. A forced harness capture now deletes the title's nvram.ram before the 5DO run, matching the reference host's .srm deletion; a fresh-NVRAM control produced the same 108/105 command cadence, ruling out saved guest settings.

1.1bh TEST-08 unique corpus and 3-way compatibility comparison (2026-11-08)

The local collection is now inventoried by content, not by filename. tools/corpus-inventory.py asks the 5DO CLI for each image's gameId (the MD5 of the first two 2048-byte sectors, the same key the library and 4DO database use) and records the database metadata. On the current J:\Panasonic 3DO plus G:\MAME\MAME Software List CHDs\3do collection:

  • 1,371 disc images (1,366 on J:, 5 on G:), 0 read errors.
  • 653 distinct game IDs — the same count as the game database.
  • 414 unique titles after collapsing identical images and grouping regional revisions.
  • 378-title game test set after separating demo, magazine/sampler, hack/translation, and audio-only images.

tools/generate-compat-comparison.py matches the unique titles against the saved 4DO wiki and the Phoenix (Arts-Union) list, then joins the 5DO sweep and manual grades. The result is docs/CompatibilityComparison.md, with the full mapping in docs/data/unique-corpus.csv and docs/data/compatibility-comparison.json:

Metric Result
Unique titles 414
Game test set 378
Matched to 4DO 232 of 241 reference titles
Matched to Phoenix 246 of 302 reference titles
Matched to both 194
5DO manual gameplay grades 107
5DO automated loadability 307
5DO untested 0
5DO ahead of 4DO 5 (Alone in the Dark, DinoPark Tycoon, Orbatak, Psychic Detective, Shootout at Old Tucson)
5DO behind 4DO 22 after Tier B batch 4 (8 initially; batches 1-4 turned unknown rows into explicit grades)
5DO behind Phoenix 25 after Tier B batch 4 (10 initially)
Loadability regressions 0

Phoenix is a 100% list, not a graded database. All 302 entries are recorded at 100%, so an unlisted title means no data. The useful comparison is the reverse direction: a title Phoenix lists as working but 5DO cannot load is a regression.

The canonical test set is written to artifacts/corpus/unique-testset.txt for tools/compat-sweep.ps1 -FileList. The 44 titles that were previously untested were run at 3,000 frames with -AllowUnsigned: 40 rendered and 4 had no video (3DO Gamepack #1, the audio-only A&M Music Sampler, Evil Australians, and the LameNES emulator). All four are secondary media/homebrew, not retail game regressions, and all 414 unique titles now have a 5DO status. This section is the evidence for TEST-08; it does not replace the 50-title accuracy panel or the playability audit.

1.1bi TEST-06 Tier B batch 1: the ten comparison diffs (2026-11-08)

The first Tier B batch covers the ten titles where 5DO differed from 4DO or Phoenix. It is a 5DO-only pass: the question is whether the shipped emulator reaches and sustains gameplay, and Opera is real-time paced. Runner: tools/playability-sweep.py; manifest: benchmarks/diff/tierb-behind.json; script: benchmarks/diff/scripts/tierb-10min.txt.

Method: 36,000 frames (10 minutes at 60 Hz), default compatible profile, panafz1 for retail discs and panafz10-patched for the arcade CD-ROMs, captures at frames 600/3,600/18,000/36,000, and a fresh per-title NVRAM. The generic script advances menus for the first minute and then rotates movement/action buttons.

Title What the captures show Verdict
Battle Chess Interactive chessboard at 3,600, 18,000, and 36,000; no crash Good (promoted)
Crash 'n Burn Black boot at 3,600; in-race HUD and car at 18,000 and 36,000 Good (promoted)
Myst Gameplay scene at 3,600, load menu at 18,000, gameplay scene at 36,000 Good (promoted)
Olympic Summer Games 100m at 3,600, discus at 18,000, high jump at 36,000 Good (promoted)
Orbatak Black boot at 3,600; arena match at 18,000 and 36,000 Good (promoted)
Crime Patrol Interactive Play/Rookie/1 Player/Calibrate menu at 18,000 and 36,000 Playable: generic pad script cannot select a lightgun target
Mad Dog II Live-action intro at 3,600 and 18,000; menu at 36,000 Playable: needs a title-specific lightgun script
Guardian War Battle scene at 3,600; save menu at 18,000; transition at 36,000 Playable: generic script does not sustain the battle
Shootout at Old Tucson In-game scene at 3,600; credits at 18,000 and 36,000 Playable: gameplay reached, then script exits to credits
Slayer Title/menu at 18,000 and 36,000; no character started Playable: needs a title-specific script

All ten runs completed with exit code 0 and dspRunning: true. The five promoted titles were re-run against a stock BIOS with no RSA patch; all produced 13,243,230 samples (18,000 fields) with non-zero audio hashes. Crash 'n Burn's 18,000-frame WAV is non-silent (peak 7,631, RMS 1,190, 0 clipped frames, 6.6% silence).

Result: the comparison moves from 8 Behind vs 4DO and 10 vs Phoenix to 4 Behind vs 4DO and 5 vs Phoenix, with 0 loadability regressions. (Batch 2 then converted seven more unknown rows into explicit Playable grades; see 1.1bj.) The remaining five rows need title-specific input scripts, not emulator fixes; that is the next Tier B increment.

1.1bj TEST-06 Tier B batch 2: 32 peer-flagged titles (2026-11-08)

Batch 2 was generated with tools/build-tierb-manifest.py: the next 32 titles where both 4DO and Phoenix claim Playable/Good and 5DO had only automated loadability evidence. The run used the same 36,000-frame generic script, --bios-auto (so the database selected the BIOS and input device), and four parallel jobs. All 32 completed with exit code 0.

25 promoted to Good: 3D Atlas, Alone in the Dark, Alone in the Dark 2, AutoBahn Tokio, Ballz - The Director's Cut, Battle Pinball, BattleSport, BC Racers, Brain Dead 13, Burning Soldier, Bust-A-Move, Cannon Fodder, Captain Quazar, Casper, Corpse Killer, Coven, Cowboy Casino, Crayon Shin-chan, Creature Shock, Cyberdillo, Cyberia, D, Daedalus Encounter, Defcon 5, and Demolition Man.

7 kept Playable:

Title Why it is not Good yet
20th Century Video Almanac Stays on the title screen; the generic script does not enter the almanac.
BladeForce Stays on the Start New Game/options menu; no mission started.
Carrier - Fortress at Sea Reaches the Carrier Battle Group info menu; gameplay not reached.
Club 3DO - Station Invasion Interactive menu only; it is a media/demo disc.
DeathKeep Shows the title, then a black screen at 18,000/36,000; needs a recheck without the generic script.
Dennis Miller - That's News To Me Reaches the Players Association show menu; it is an FMV comedy disc.
Digital Dreamware Reaches the music visualizer/title; it is a media disc, not a game.

An 18,000-frame --audio-hash pass covered all 32 titles: every run produced 13,243,230 samples (Defcon 5: 15,876,000, its PAL/field timing), a non-zero audio hash, and dspRunning: true. The ledger is docs/data/playability-grades.csv.

Result: 43 titles now have a manual grade (30 Good, 13 Playable). The Behind count is now 10 vs 4DO / 12 vs Phoenix, but that is better evidence, not a regression: seven new Playable grades replaced unknown loadability rows, and every one has a recorded reason. Batches 1-2 closed the unknown-review gap; the remaining work is the seven title-specific scripts above plus the five batch-1 titles.

1.1bk TEST-06 Tier B batch 3: 32 more peer-flagged titles (2026-11-08)

Batch 3 used the same generator and 36,000-frame --bios-auto pass. All 32 completed with exit code 0.

24 promoted to Good: DinoPark Tycoon, Doctor Hauzer, Doom, Dragon Lore, Dragon Tycoon Edge, Dragon's Lair, Drug Wars, Escape from Monster Manor, FIFA International Soccer, Flashback, Flying Nightmares, Foes of Ali, Gex, Gridders, Horde, Icebreaker, Incredible Machine, Iron Angel of the Apocalypse, Iron Angel Returns, Jammit, Jikki Pachi-Slot Simulator Vol. 1, Johnny Bazookatone, Last Bounty Hunter, and Lemmings (Lemmings keeps its existing Fun 4 passcode gameplay evidence; the generic pass stops at the title menu).

8 kept Playable:

Title Why it is not Good yet
ESPN Golf - Shot Making Instructor video/lesson, not a playable round; instructional disc.
Family Feud Stays on the Music/Speech/Sound/Rapid Play/Difficulty options menu.
Hell - A Cyberpunk Thriller Stays on the title menu.
Insect War Stays on the character-select screen.
John Madden Football Stays on the Scouting Report/play-selection menu.
Jurassic Park Interactive Reaches the Operations terminal map; in-park gameplay not confirmed.
Kingdom - The Far Reaches Stays on the title menu.
Life Stage - Virtual House Sample-menu/house content; it is a virtual-house disc.

An 18,000-frame audio-hash pass covered all 32 titles with non-zero hashes and dspRunning: true.

Result: 75 titles now have a manual grade (54 Good, 21 Playable) and 339 titles still have automated loadability only. The Behind count is 17 vs 4DO / 20 vs Phoenix, again because eight more unknown rows became explicit Playable grades. The review gap is closing; the grade gap is now the real work.

1.1bl TEST-06 Tier B batch 4: the last 32 peer-flagged titles (2026-11-08)

Batch 4 was generated with tools/build-tierb-manifest.py and covers the last 32 titles where both 4DO and Phoenix claim Playable/Good and 5DO had only automated loadability evidence. The run used the same 36,000-frame generic script, --bios-auto, and four parallel jobs. All 32 completed with exit code 0 and dspRunning: true.

27 promoted to Good: Lost Eden, Lost Files of Sherlock Holmes, Lucienne's Quest, Mad Dog McCree, Mazer, MegaRace, Microcosm, Need for Speed, Novastorm, Out of This World, OverDrivin', PaTaank, Pebble Beach Golf Links, Perfect General, Phoenix 3, Plumbers Don't Wear Ties, PO'ed, Policenauts, Primal Rage, Psychic Detective, Quarantine, Real Pinball, Return Fire, Rise of the Robots, Road Rash, Panzer General, and Quarterback Attack.

5 kept Playable:

Title Why it is not Good yet
N. O. B. - Neo Organic Bioform Stays on the Mode Select menu; the generic script does not start Survive/Versus.
Off-world Interceptor Stays in the vehicle upgrade/shop menu; no mission launched.
Olympic Soccer Stays on the team-select and title menu; no match started.
PGA Tour 96 Stays on the Main Menu; the generic script does not start a round.
Return Fire - Maps O' Death Stays on the expansion title/box-art screen; it is an add-on disc.

Three follow-up checks changed or confirmed grades:

  • Panzer General was Playable in the generic pass. A title-specific benchmarks/diff/scripts/tierb-panzer-start.txt script confirmed START GAME with the A button and reached the scenario map with units at 9,000 frames, so it is Good.
  • Quarterback Attack was Playable in the generic pass. A title-specific benchmarks/diff/scripts/tierb-quarterback-start.txt script moved from Skill Level to KICKOFF, confirmed it, and reached football gameplay (scoreboard and play select) at 20,800 frames, so it is Good.
  • Real Pinball showed a nearly black late frame in the generic pass. A five-capture recheck (artifacts/tierb-panel-4-realpinball-check) showed the full table at frame 30,000; the dark frame is the attract/transition view, not a render failure. It stays Good.

An 18,000-frame --audio-hash pass covered all 32 titles: every run produced 13,243,230 samples, a non-zero audio hash, and dspRunning: true (artifacts/tierb-panel-4-audio/summary.csv). The ledger is docs/data/playability-grades.csv.

Result: 107 titles now have a manual grade (81 Good, 26 Playable) and 307 titles still have automated loadability only. The Behind count is 22 vs 4DO / 25 vs Phoenix; all of the new rows are explicit Playable grades with a recorded reason or a missing title-specific script, not loadability regressions. Five titles are now ahead of 4DO. The ~100-title Tier B target is met; converting the remaining Playable menus into confirmed gameplay is the next grade-gap increment.

1.1bm TEST-06 Tier C: first gameplay-depth increment (2026-11-09)

Tier C targets the 25 titles still graded Playable after Tier B. The Tier B generic script presses A early and often stops before the menu appears; the 3DO C button is the usual confirm/select button in these menus. Added:

  • benchmarks/diff/tierc-panel.json - the 25-title remaining-Playable gap.
  • benchmarks/diff/tierc-batch-3.json and tierc-batch-4.json - the C-button and follow-up batches.
  • benchmarks/diff/scripts/tierc/ - title-specific start and gameplay scripts.
  • artifacts/tierc-baseline - a 25-title baseline with the generic Tier B script; 25/25 completed with exit code 0 and all captures written.

The scripts use C for menu confirmation, press it at the frame where the menu is actually visible, and switch to A/B/directions for gameplay. A Good grade requires reaching gameplay and completing the checkpoint window without a hang or crash.

Promoted to Good in this increment (batch tierc-1):

Title Evidence
Olympic Soccer tierc/olympic-soccer-play.txt: ARCADE -> team select -> FULL-TIME at 18,000 frames; returns to the menu by 36,000.
John Madden Football tierc/john-madden-c2.txt: C through the Scouting Report and play selection; live football field, players, and clock at 12,000/24,000/36,000 frames.
Crime Patrol tierc/crime-patrol-c2.txt: C selects PLAY and 1 PLAYER; live-action lightgun scene at 48,000 frames.
Mad Dog II - The Lost Gold tierc/mad-dog-ii-c2.txt: C selects Play and 1 Player; live-action stage at 48,000 frames.
Shootout at Old Tucson tierc/shootout-c2.txt: skips the credits with C; live-action lightgun scene with the ammo HUD at 48,000 frames.
Guardian War tierc/guardian-war-play3.txt: keeps the game in play after the opening battle; Goddess dialogue at 18,000 and the world-map character at 24,000/36,000 frames.
Hell - A Cyberpunk Thriller tierc/hell-play6.txt: A activates the menu, three Downs select Start New Game, and A confirms; character select at 40,000 and 3D adventure gameplay at 44,000/48,000 frames.
Slayer tierc/slayer-play8.txt: Create Character, Choose Face, and Dungeon Settings confirmed; first-person dungeon gameplay with map and HP/FD HUD at 80,000/90,000 frames.

The CSV ledger now has 101 rows: 84 Good, 17 Playable. The remaining Playable rows are five media/reference discs (Club 3DO, Dennis Miller, Digital Dreamware, ESPN Golf, Life Stage), the Return Fire - Maps O' Death data disc, and eleven game/script gaps. The strongest remaining leads are BladeForce (3D mission/cinematic scenes at 24,000/36,000; interactive gameplay not yet confirmed), PGA Tour 96 (Player List, Pro Info, and Golf Bag; bag confirm not scripted), N.O.B. (SURVIVE scene at 18,000), Insect War (3D content at 30,000), DeathKeep (late title menu), Kingdom (late title menu), Family Feud (two-stage game options), and Jurassic Park (play-level screen). Carrier's briefing room appears to be pointer-driven, so the scripted pad cannot select Challenge!; it stays Playable with that reason.

Fixed-script recheck (2026-11-09): an early batch generator emitted a press and its None release at the same frame. InputScript.Parse coalesces duplicate frames, so the press could be discarded. Every affected script was repaired with an explicit 20-frame hold and the affected titles were re-run in artifacts/tierc-fixed, tierc-batch-7, and tierc-batch-8. The corrected runs confirmed John Madden, Crime Patrol, Mad Dog II, Shootout at Old Tucson, Guardian War, Hell, and Slayer, and withdrew the PGA Tour 96 promotion because the fixed script stays on the Player List. The paired BladeForce input/no-input probe now shows a real input effect, but both runs only expose live-action cutscenes and return to the menu; BladeForce remains Playable.

The Tier C workflow and grade rules are in CompatibilitySweep.md 4.3.

1.1bn TEST-06 Tier C: remaining-lead probes (2026-11-09)

Batches 9-12 targeted the last Playable leads with longer windows, alternate confirm buttons, a mouse-device probe, and paired input/no-input runs:

  • PGA Tour 96. The C-driven path reaches the Player List, adds a pro (Stadler/Tom Kite/Brad Faxon), opens Pro Info, and reaches the Golf Bag screen. Confirming the bag cycles back to Pro Info instead of starting a round. Evidence: artifacts/tierc-batch-10, tierc-batch-11, tierc-batch-12.
  • Family Feud. The options screen is reached, but Down/C toggles SPEECH and RAPID PLAY values and does not reliably reach DONE or the Game Options screen. Evidence: artifacts/tierc-batch-9/family-feud-play6.
  • DeathKeep. Selecting Start Game plays the stained-glass intro and then a black screen at 44,000/48,000; no gameplay. Evidence: artifacts/tierc-batch-9/deathkeep-play7.
  • Kingdom - The Far Reaches. A-only and repeated A/C title input start the intro and return to the title by 60,000/90,000; no gameplay. Evidence: artifacts/tierc-batch-9/kingdom-play5.
  • BladeForce. The 60,000-frame input run still only exposes live-action cutscenes and returns to Main Options; the paired no-input run stays on the menu. Remains Playable as FMV-heavy content. Evidence: artifacts/tierc-batch-5/bladeforce-*, artifacts/tierc-batch-9/bladeforce-play4.
  • Carrier - Fortress at Sea. The briefing room is unchanged with both the pad and the 1=mouse device; the pointer-driven UI cannot be scripted with the current source. Evidence: artifacts/tierc-batch-9/carrier-mouse.

No grade changes came from these probes; the ledger stays 101 rows: 84 Good, 17 Playable.

1.1bo TEST-06 Tier C: second gameplay-depth increment (2026-11-09)

The second increment added pointer state to the input-script format and re-checked the menu leads whose earlier scripts pressed buttons before the menu was visible.

Input-script pointer support. InputScript now accepts one optional pointer token per line after the pad buttons: @mouse dx dy [left,middle,right,shift], @lightgun x y [trigger,option,reload,offscreen], @arcadegun x y [trigger,service,coins,start,holster,offscreen], and @trackball dx dy [start1,start2,coin1,coin2,service]. ScriptedPbusSource applies the state to every port configured with that device, keeps buttons held, and treats relative mouse/trackball motion as a one-frame impulse. The parser and source are covered by six new InputScriptTests. The reference OperaCore bridge is still joypad-only, so pointer scripts are 5DO-only evidence.

Promoted to Good (tierc-2):

Title Evidence
Carrier - Fortress at Sea tierc/carrier-play5.txt: the briefing room's pad path opens the trap menu (DAY/NIGHT/WEATHER TRAPS, PREFLIGHT BRIEFING), then starts a live trap approach with the 3D ocean, aircraft HUD, and a WAVE OFF glide-slope message at 8,000/14,000 frames. The earlier probe pressed buttons before the briefing room appeared, which is why it looked pointer-only.
Family Feud tierc/family-feud-play7.txt: eight Down presses move from the options rows to DONE; the game-mode menu opens and FAMILY VS 3DO SYSTEM starts the show. Live questions, host, contestants, the $2,000/$3,000 bank, and the letter keyboard render at 9,000/12,000 frames.
Insect War tierc/insect-war-play5.txt: Story Mode, character confirm, and match start reach the first-person HUD (timer, health bars, radar, crosshair) at 58,000/62,000 frames; the WARNING/explosion state follows. No crash through 72,000.
N. O. B. - Neo Organic Bioform tierc/nob-c2.txt: SURVIVE starts and reaches the first-person corridor with the VITALITY HUD at 9,000 frames. The session later returns to MODE SELECT, and the script re-enters. No crash through 36,000.

The four promotions and the BladeForce path are reproducible from benchmarks/diff/tierc-batch-13.json:

python tools/playability-sweep.py --manifest benchmarks/diff/tierc-batch-13.json `
  --out artifacts/tierc-batch-13 --jobs 4

The ledger is now 101 rows: 88 Good, 13 Playable. Seven of the remaining Playable rows are media/reference/add-on discs (20th Century Video Almanac, Club 3DO, Dennis Miller, Digital Dreamware, ESPN Golf, Life Stage, and the Return Fire Maps O' Death data disc). The six remaining game leads are BladeForce, DeathKeep, Jurassic Park, Kingdom, Off-world Interceptor, and PGA Tour 96. Jurassic Park's Operations hub shows the pause menu's TIME LEFT/LIVES LEFT session state, but no in-park mini-game was opened; PGA Tour 96 still cycles Golf Bag -> Pro Info, and Off-world Interceptor remains in the pre-mission shop.

DeathKeep differential (2026-11-09). The Tier C black-screen lead is not a 5DO defect: the reference OperaCore reached 48,000 frames with the same script and its f048000.bmp capture is also pure black, while 5DO's f048000.png is black at the same checkpoint. The only comparison field that diverged at 48,000 was the sample count (35,315,280 vs 35,315,323, the documented 1.214 ppm reference-clock alignment). The Start Game path therefore leads to a legitimate long black screen/load on both engines, not a 5DO renderer or CD regression. Evidence: artifacts/diff/tierc-triage2/deathkeep-play7.

BladeForce differential (2026-11-09). The reference core reaches the same Main Options menu at 60,000 frames; the only comparison difference is the documented sample count (44,144,100 vs 44,144,154). A new script (tierc/bladeforce-play6.txt) moves Right to START NEW GAME, confirms the name entry, and plays the live-action intro and character-platform cutscenes through 100,000 frames. The menu/cutscene loop is therefore a content/script limit, not a 5DO input or renderer bug; BladeForce stays Playable because no interactive mission was reached. Evidence: artifacts/bladeforce-play6; artifacts/diff/tierc-triage2/bladeforce-play4.

Kingdom differential (2026-11-09). A 60,000-frame comparison (artifacts/diff/kingdom-triage-60k) captures the same Kingdom - The Far Reaches title screen on 5DO and OperaCore. The A/C intro loop is therefore a content/script limit on both engines, not a 5DO boot or renderer defect; Kingdom stays Playable because this script starts no scenario.