ARQ over A-MPDU/BlockAck: quantify ACK-asymmetry reduction and goodput on the arq_e2e bench

[josephnef]

Context (measured, #359)

Per-frame ACK has a small measured asymmetry: ~15–20 duplicate deliveries per phase (frame arrived, ACK lost, peer retransmitted) and 1–3 ok=0-but-delivered per phase (peer wrote off a delivered frame). Safe direction — the TX ledger under-claims — but it wastes airtime on redundant retransmissions, and it scales with ACK exchanges. BlockAck replaces N ACK exchanges with one re-solicitable bitmap, and A-MPDU is already worth ~+30% goodput (docs/aggregation.md); the responder side is already BA-capable (SetAckResponder arms the BlockAck responder too).

Steps

1. [verify] Report semantics under aggregation. Vendor check for J3: does CCX report per MPDU or per PPDU inside an A-MPDU, and how does the SW_DEFINE tag echo behave? The arq_e2e tag-reconstruction depends on the answer. (The payload pctr ledger is immune — it is per-MPDU by construction; the A-MPDU engine renumbers 802.11 seqs, which is why the bench never keyed on them.)
2. [measure] Bench arm. tests/arq_e2e_delivery.sh with DEVOURER_TX_AMPDU_MODE on the drone-sim + BA responder on the DUT: acked_undelivered (must stay 0), dup rate, ok=0-but-delivered, goodput vs the per-frame-ACK arms — same phases, same stalls.
3. [revisit] Retry interplay. The earlier aggregation work found the data-queue aggregation path ignores noack and rty0 suppresses — re-examine under the DEVOURER_TX_RETRY_LIMIT knob so the BA path's retransmission budget is actually configurable.
4. [document] If the contract holds and asymmetry shrinks: recommend A-MPDU+BA as the ARQ transport for video streams in docs/aggregation.md + docs/scheduled-mac.md, with the measured numbers.

Acceptance

au = 0 under BA in all arms; dup + write-off rates compared against per-frame ACK with the same RF; goodput delta published.

Refs: #359, docs/aggregation.md, the SetAmpduMode/BA-responder work.

[josephnef]

Step 1 vendor check done: CONDITIONAL GO. CCX reports are per-MPDU with per-MPDU SW_DEFINE echo — aggregation does not change report granularity (rtl8852 ccxrpt struct carries per-MPDU tx counts; the 8822C ops stamp SW_DEFINE per descriptor and descriptors map 1:1 to MPDUs). The vendor only stamps SW_DEFINE on its ack_report/mgmt path, but that is a driver choice — devourer stamps every data frame itself, so tag correlation carries over. Two bench-arm redesign requirements before running: (1) BA-driven retransmits can emit MULTIPLE reports for the same tag with incrementing data_tx_cnt — the analyzer's tag-unwrap assumes one report per frame, so it needs a per-tag reconciliation rule (last-report-wins for state, max for retries); (2) monitor MISSED_RPT_NUM-style pressure — dense aggregates multiply the report demand against the ~1.3k/s emission ceiling, so the arm should run with DEVOURER_TX_REPORT sampling from the start.

[josephnef]

Measured — full write-up in #373 (closes this on merge). Headline per acceptance item:

1. Contract: acked_undelivered=0 in every BA arm — ACK stays admission-gated under AGG_EN, no phantom ACKs.
2. Telemetry: per-frame CCX does not survive aggregation — 19,516 reports vs the control's 49,636 over the same span (−61%, suppressed in whole-aggregate bunches), retries=0 on every report (BA retransmit cycles invisible), zero multi-report frames. Tag-gap accounting is un-aggregated-only; receipts (tx.report at rate: attribute the coverage collapse (53% @ ~2.3k fps), deterministic SPE_RPT sampling, windowed RX receipts #363) are the delivery truth under A-MPDU.
3. Goodput at 4 k fps saturation: BA −8% delivered vs the identical un-aggregated feed (208,792 vs 226,691) at MCS3/512 B — the +30% docs figure is high-MCS broadcast, and both numbers are now paired in docs/aggregation.md.

Recommendation shipped: leave A-MPDU off on the ARQ/FPV shape; if revisited at high MCS, use receipts-only accounting (the analyzer's per-tag reconciliation is ready for a multi-report future).