VolkerChristian · Sep 30, 2026 · Sep 30, 2026
diff --git a/docs/USER_GUIDE.md b/docs/USER_GUIDE.md
diff --git a/docs/design/SCOPED_SETTINGS_IMPLEMENTATION.md b/docs/design/SCOPED_SETTINGS_IMPLEMENTATION.md
@@ -1,145 +1,69 @@
-# Flat settings pages and three-level inheritance
+# Scoped settings: schema, drafts and Preferences
 
-Requested 27 September 2026. This supersedes the nested navigation introduced by
-the canonical Settings consolidation. The user approved up to 500 net additional production lines for this change.
-Implementation and Linux verification are recorded in [VERIFICATION.md](../VERIFICATION.md).
+The October 1 refinement supersedes the September 27 combined application dialog.
+See [verification and accounting](../reviews/2026-10-01-settings-refactor.md).
 
 ## Invariant
 
-For an inheritable setting, the effective value is the explicit slide override,
-otherwise the explicit talk override, otherwise the application default. The same
-resolved values must drive presentation, voice summaries, preview, narration,
-speech synthesis and audio validity. Missing overrides are different from zero,
-empty text, false and an explicitly selected model default.
+Project resolves slide override → talk setting → application default → built-in.
+Resolution, schema v5, SETTING_DEFAULTS, language versions and audio fingerprints
+are unchanged. MainWindow owns the live project. A scoped dialog owns a draft,
+never the live project. Preferences owns computer controls, not talk defaults.
 
-Three fixed-scope Settings dialogs share flat navigation with up to five pages:
+## Ownership
 
-1. Voice & language: voice identity/source, language, writing style and delivery.
-2. Talk & preparation: audience/conference and timing policies; a read-only talk mode row.
-3. Presentation & recording: playback behaviour, recording, output and added audio.
-4. AI & speech engine: Codex selection, Qwen options and GPU lifecycle.
-5. Application: account, device preferences and general application settings.
-
-Pages combine related settings. Application defaults, talk and slide each have a
-fixed-scope window; there is no scope selector. Navigation exposes only applicable
-pages. Device and engine lifetime retain application scope. Existing shortcuts
-select the concrete page and focus target; native style and palette are retained.
-
-## Settings ownership
-
-| Scope chain | Settings | Editor location |
+| Scope | Settings | Editor location |
 | --- | --- | --- |
-| Application → talk → slide | Voice, writing style, delivery attributes/directions | Corresponding scope's Voice & language dialog |
-| Application → talk → slide | Narration language | Application default / main Language combo / slide Language of this slide; talk dialog readout + reset only |
-| Application → talk → slide | After-slide action, pause | Application and talk Playback sections; slide inspector After this slide / Timing |
-| Application → talk | Mode; recording intent | Application default dialog; main toolbar combo / main checkbox; talk dialog readouts + reset only |
-| Application → talk | Language arrangement, Codex model/reasoning, synthesis sampling, Quick/Realtime policies, tolerance, recording source/microphone/policy, encoding, background gain/loop | Corresponding scope's task page |
-| Application only | Account, GPU lifetime, device preferences, libraries | Application dialog and focused menu entrances |
-| Talk only | Duration | Main Duration spin only |
-| Talk only | Title, conference/sources, audience/objective, background track, output destination, gradual delivery | Talk dialog; background track controls share Background audio |
-| Talk only | Language versions | Main Language combo |
-| Slide only | Narration, notes, inclusion, budget, additional clips | Editor, inspector Timing/inclusion, slide Additional audio dialog |
-
-
-The selected language version belongs to talk scope; it is not a fourth tier.
-Slide overrides belong to that version's slide. Explicit passage language markers
-remain content instructions and must satisfy the talk's language-arrangement rule.
-Changing language does not translate existing text automatically.
-
-Voice identity is an atomic selection, including its reference recordings. Vocal
-attributes inherit individually. Supported controls depend on the effective voice;
-the Qwen variant follows that voice rather than an independently conflicting choice.
-
-## Concrete implementation boundaries
-
-1. **Project/settings state — `project.py`, application preferences.** Introduce
-   one allowlisted settings resolver and explicit override storage. Application
-   defaults are stored in application config; talk/slide overrides in the talk
-   JSON. Workers receive a resolved snapshot instead of reading live preferences.
-   Replace direct effective-setting calculations at consumers rather than keeping
-   a parallel policy. Migrate existing talk values to explicit overrides, preserving
-   saved behaviour and existing audio when effective synthesis inputs are unchanged.
-
-2. **Bindings and dialog — `ui.py`, `options.py`, `app.py`.** Replace the tree and
-   thirteen pages with five combined pages. Reuse the field editors with explicit
-   scope-aware bindings, inherited-value captions and removal of overrides. Keep
-   ordinary controls direct; technical parameters remain under Advanced. Preserve
-   pending edits while changing pages. Each fixed-scope dialog has its own
-   Save/Cancel transaction. Immediate account,
-   engine, system-audio and explicit library operations stay clearly identified.
-   Do not temporarily adopt a fake talk to edit application defaults.
-
-3. **Voice identity — `project.py`, `voices.py`, voice actions in `app.py`.** Preserve
-   recorded/imported/designed/predefined origin independently of the synthesis
-   backend and user-assigned name. Accepting a design keeps its origin after its
-   audio becomes a Base reference. Saved-library selection is not another voice
-   type. Older ambiguous references are labelled Reference voice without guessing.
-   Use the resolved identity in both talk and selected-slide summaries. Reuse the
-   existing reference-copy/hash mechanism for portable talk assets.
-
-4. **Generation and engine ownership — `services.py`, `runtime.py`.** Resolve
-   language, writing and speech settings per slide. Preserve whole-talk context
-   during narration generation. Replace the one-configuration-per-batch assumption
-   with compatible batches through the existing single speech owner, including
-   Realtime and explicit preload. Avoid holding the existing session guard across
-   a transition that needs to acquire it again. Speaker changes within a compatible
-   model should not trigger needless model reloads. Model-type changes may require
-   loading and must report that work. Do not change the existing text splitting.
-
-5. **Modes, cache and saved projects.** Prepared, Quick and Realtime share effective
-   settings; remove Quick's hidden Ryan/Professional substitutions. Quick keeps
-   its simplified start workflow. Audio fingerprints use effective synthesis inputs,
-   not override provenance, so switching from inherited to explicit identical values
-   preserves valid audio. Only affected slides become outdated. Keep existing words
-   and saved files until explicit replacement succeeds. Playback and exports use the
-   accepted artifacts, not newly resolved settings midway through an operation.
-
-## Defaults and persistence semantics
-
-Inheritance is live: changing an application default affects talks/slides that still
-inherit that field. Show the affected scope when saving application defaults; do
-not silently regenerate text or audio. An active job retains its starting snapshot.
-
-Provide an explicit Keep these settings for this talk action to turn inherited
-values into talk overrides and include required voice assets. This makes a talk
-independent of another computer's defaults without another inheritance tier.
-Account credentials and machine-specific choices are never copied into a talk.
-
-## Verification
-
-- Resolver precedence, removal of overrides, explicit empty/zero/false values and
-  supported/unsupported vocal controls; effective-value equality preserves cache.
-- Application config versus talk JSON persistence; migration and reopen with changed
-  application defaults; pinning settings and copying the talk to another directory.
-- Predefined, own and designed voices retain their displayed origin through preview,
-  acceptance, library save/load and project reopen; library audition does not select.
-- Three flat fixed-scope dialogs, five possible pages, focused shortcuts, scoped Save/Cancel,
-  application defaults without a talk and correct selected-slide/version binding.
-- Whole-talk and per-slide text/audio generation, mixed-language validation and
-  Prepared/Quick/Realtime use the same resolved settings; no text approval is added.
-- Engine reuse and compatible/incompatible voice transitions, cancellation and
-  progress; pre-existing valid audio avoids loading a model unnecessarily.
-- Linux native Qt/Breeze Xvfb, minimum/default window sizes and 150% scaling; full
-  regression suite, packaged launcher and library audit. Only Linux is tested.
-
-## Accounting and approval
-
-The prior approved limit was 160 net additional production lines from baseline
-117fb8b. Before this pass, the tree used 101, leaving 59. Flattening navigation can reduce UI
-code, but the existing model has no persisted application speech/language defaults
-or general slide overrides, and generation is bound to talk-level speech config.
-Moving widgets cannot supply those missing semantics.
-
-Estimated additional production growth, after removing superseded paths:
-
-| Area | Net lines |
-| --- | ---: |
-| Settings resolution, serialization and migration | 110–150 |
-| Scope-aware editors and transaction changes, offset by navigation consolidation | 130–180 |
-| Voice origin, summaries and library/reference persistence | 30–50 |
-| Generation, mode and single-engine transition changes | 80–120 |
-| **Total** | **350–500** |
-
-Approved allowance: at most **500 net additional production lines from the current
-tree** (601 total from 117fb8b). Tests and documentation are counted separately.
-This is an estimate and ceiling, not a target. Final measured accounting is recorded with the verification results.
+| Application → talk → slide | Voice identity and references | Voice & language in each fixed-scope dialog |
+| Application → talk → slide | Language | Talk defaults; main Language combo; Slide sound → Voice & language |
+| Application → talk → slide | Writing style, delivery instructions/attributes | Writing & delivery in each fixed-scope dialog |
+| Application → talk → slide | After-slide action and pause | Defaults/talk Timing & playback; slide inspector only |
+| Application → talk | Mode | Talk defaults Timing & playback; main Mode combo |
+| Application → talk | Recording intent | Talk defaults Audio & recording; main checkbox |
+| Application → talk | Language arrangement | Defaults/talk Voice & language |
+| Application → talk | Quick/Realtime/timing policies | Defaults/talk Timing & playback |
+| Application → talk | Codex model/effort; synthesis sampling | Defaults/talk AI model |
+| Application → talk | Recording/output settings; background gain/loop | Defaults/talk Audio & recording |
+| Talk | Duration | Main Duration spin only |
+| Talk | Title/audience/objective/conference/sources; pin inherited settings | Talk settings → Talk |
+| Talk | Background file and recording destination | Talk settings → Audio & recording |
+| Slide | Inclusion/budget | Slide flow inspector |
+| Slide | Audio clips, gain and placement | Slide sound → Audio & recording |
+| Computer | Display, system sound, account, engine and lifecycle policy | Preferences, immediate; Close only |
+| Shared library | Reusable voices and delivery presets | Relevant component's explicit library actions, immediate |
+
+Talk mode, language and recording rows are read-only with Reset. Every inheritable
+row has a constant source caption (set here/from talk/from app/built-in) and Reset;
+explicit zero, false, empty, and equal-to-parent values remain overrides.
+
+## Implementation
+
+`settings_schema.py` is the presentation metadata authority for keys, kinds,
+choices, scope, page, section and order. Values/scopes originate in Project's
+SETTING_DEFAULTS and SETTING_CHOICES. The explicitly listed custom owners handle
+voice, sampling, delivery presets and assets. `SettingField` builds and loads the
+same editor/source/Reset row in dialogs and the inspector. Widgets are placed once.
+
+`ScopedSettingsDialog` has thin TalkDefaults/Talk/Slide configurations. It composes
+VoiceEditor, AudioAssets, CodexModelPicker, ConferencePanel and DeliveryPresets.
+The narrow controller API is documented in module docstrings and checked by AST.
+PreferencesDialog owns device, account and engine widgets. MainWindow owns actions,
+the speech session and persistence; no compatibility widget aliases remain.
+
+Opening copies the project (or defaults) and records its generation. Edits use
+`apply_edit`, shared by draft editing and Save replay, and append immutable operations.
+Save checks the generation, replays once through Project mutators inside the
+existing configuration change boundary, updates audio/UI once, compares stale
+slides, and writes once. Defaults write QSettings; talk/slide write the manifest.
+Cancel writes nothing, preserves pre-existing main-window modifications, cancels
+owned operations and deletes only session-created unreferenced media.
+
+Conference/import workers receive the draft and never save. Late cancelled results
+cannot edit a reopened draft. Save is unavailable during project-returning jobs;
+Cancel remains available. Shared-library actions do not save the project.
+Preferences account/model jobs do not save pending document settings.
+
+Draft edits never call MainWindow.edit_setting, configuration_changed or refresh.
+MainWindow no longer reloads drafts after live edits. Local timing is measured on
+the draft when its page is shown/edited. Engine status and account availability are
+read-only external state; they do not replace draft model selections.
diff --git a/docs/design/UX_PLACEMENT_PROPOSAL.md b/docs/design/UX_PLACEMENT_PROPOSAL.md
@@ -8,6 +8,44 @@ interaction contract. The [original inventory](UI_INVENTORY.md) is the historica
 baseline. Existing visual styling is retained. The implementation was authorized
 with a ceiling of 160 net additional production lines.
 
+## Current settings and access contract — October 1
+
+The [scoped settings contract](SCOPED_SETTINGS_IMPLEMENTATION.md) supersedes the
+historical settings layout below. The main controls and single-owner rules remain.
+Preferences is this computer (immediate, Close). Talk defaults, Talk settings and
+Slide sound edit independent drafts (Save/Cancel). No scope selector exists.
+
+| Controls | Current entrance / location | Priority / navigation depth |
+| --- | --- | --- |
+| Mode, duration, language, record, Start, text/audio/play, slide navigation | Main window | P1 / 0 |
+| Predefined voice and sample | Main voice button → Voice & language | P1 / 1 |
+| Personal/designed/saved voice | Main voice button → source tab | P2 / 2 |
+| Audience, objective, conference | Talk settings → Talk | P2 / 1 |
+| Delivery style and presets | Voice button → Writing & delivery | P2 / 2 |
+| Vocal attributes | Writing & delivery → More vocal attributes | P3 / 3 |
+| Slide include/after | Slide flow inspector | P3 / 0 |
+| Slide budget/pause, inheritance reset | Slide flow → Timing | P3 / 1 |
+| Slide voice/language | Slide sound… → Voice & language | P3 / 1 |
+| Slide directions/attributes | Slide sound… → Writing & delivery [→ More vocal attributes] | P3 / 2–3 |
+| Slide clips | Additional audio… → Audio & recording | P3 / 1 |
+| Recording source, microphone, destination, background | Presentation settings → Audio & recording | P2 / 1 plus scroll if needed |
+| Pause, tolerance, Quick/Realtime policies | Talk settings → Timing & playback | P2–P3 / 2 |
+| Fit duration | Talk → Fit duration…; or talk Timing → Save and fit… | P2 / 1–2 |
+| Display/system sound | Settings → Preferences… → Display & sound (Ctrl+,) | P2 / 2; shortcut opens directly |
+| Account | Settings → Account… | P3 / 2 |
+| Speech engine | Status-bar engine button → Preferences/Speech engine | P3–P4 / 1 |
+| Inheritable application defaults | Settings → Talk defaults… → relevant page | P3 / 2–3 |
+| Codex model/effort and speech sampling | Talk settings → AI model | P4 / 2 plus scroll if needed |
+
+Count navigation separately from entering values, executing actions and Save.
+Pages use the same order at each scope, omitting empty pages. P1 main controls
+remain directly visible at 940×680. Fixed-scope titles, source captions and Reset
+identify ownership; Cancel does not save or undo unrelated changes. See the
+[October 1 evidence](../reviews/2026-10-01-settings-refactor.md) for sizes, screenshots,
+keyboard/pointer tests, test counts and measured edit/Save latency.
+
+## Historical placement records
+
 ## Current settings contract
 
 The subsequent [flat scoped-settings contract](SCOPED_SETTINGS_IMPLEMENTATION.md)

diff --git a/docs/design/UX_PLACEMENT_VERIFICATION.md b/docs/design/UX_PLACEMENT_VERIFICATION.md
@@ -5,6 +5,44 @@ Implemented 27 September 2026 against the accepted
 [inventory](UI_INVENTORY.md) remains the pre-refactor coverage baseline; the
 [user guide](../USER_GUIDE.md) describes the current interaction routes.
 
+## Current settings and access contract — October 1
+
+The [scoped settings contract](SCOPED_SETTINGS_IMPLEMENTATION.md) supersedes the
+historical settings layout below. The main controls and single-owner rules remain.
+Preferences is this computer (immediate, Close). Talk defaults, Talk settings and
+Slide sound edit independent drafts (Save/Cancel). No scope selector exists.
+
+| Controls | Current entrance / location | Priority / navigation depth |
+| --- | --- | --- |
+| Mode, duration, language, record, Start, text/audio/play, slide navigation | Main window | P1 / 0 |
+| Predefined voice and sample | Main voice button → Voice & language | P1 / 1 |
+| Personal/designed/saved voice | Main voice button → source tab | P2 / 2 |
+| Audience, objective, conference | Talk settings → Talk | P2 / 1 |
+| Delivery style and presets | Voice button → Writing & delivery | P2 / 2 |
+| Vocal attributes | Writing & delivery → More vocal attributes | P3 / 3 |
+| Slide include/after | Slide flow inspector | P3 / 0 |
+| Slide budget/pause, inheritance reset | Slide flow → Timing | P3 / 1 |
+| Slide voice/language | Slide sound… → Voice & language | P3 / 1 |
+| Slide directions/attributes | Slide sound… → Writing & delivery [→ More vocal attributes] | P3 / 2–3 |
+| Slide clips | Additional audio… → Audio & recording | P3 / 1 |
+| Recording source, microphone, destination, background | Presentation settings → Audio & recording | P2 / 1 plus scroll if needed |
+| Pause, tolerance, Quick/Realtime policies | Talk settings → Timing & playback | P2–P3 / 2 |
+| Fit duration | Talk → Fit duration…; or talk Timing → Save and fit… | P2 / 1–2 |
+| Display/system sound | Settings → Preferences… → Display & sound (Ctrl+,) | P2 / 2; shortcut opens directly |
+| Account | Settings → Account… | P3 / 2 |
+| Speech engine | Status-bar engine button → Preferences/Speech engine | P3–P4 / 1 |
+| Inheritable application defaults | Settings → Talk defaults… → relevant page | P3 / 2–3 |
+| Codex model/effort and speech sampling | Talk settings → AI model | P4 / 2 plus scroll if needed |
+
+Count navigation separately from entering values, executing actions and Save.
+Pages use the same order at each scope, omitting empty pages. P1 main controls
+remain directly visible at 940×680. Fixed-scope titles, source captions and Reset
+identify ownership; Cancel does not save or undo unrelated changes. See the
+[October 1 evidence](../reviews/2026-10-01-settings-refactor.md) for sizes, screenshots,
+keyboard/pointer tests, test counts and measured edit/Save latency.
+
+## Historical placement records
+
 ## Subsequent settings refinement
 
 The settings-specific locations below describe the preceding canonical-dialog