feat(SCOPONE-0010): improve capture pacing and settings

docs/FINDINGS.md

@@ -1,34 +1,40 @@
 # Findings
 
-> Last Updated: 2026-04-09T00:00:00.000Z
+> Last Updated: 2026-04-09T20:59:51.000Z
 
 ## Summary
 
-Initializer refresh for SCOPONE-0009. The cached findings were stale relative to the live source tree, so the observations below reflect the current Phaser, worker, and AI implementation.
+Initializer refresh for SCOPONE-0010. The cache was invalid because `docs/FINDINGS.md` no longer matched its recorded hash, and the architecture document no longer matched the live source layout after settings, preferences, and benchmark changes. The observations below reflect the current repository state.
 
 ## Codebase Observations
 
-- Primary gameplay code currently lives in 10 TypeScript source files under `src/`; the Android wrapper adds 3 Java files.
+- Primary gameplay code currently lives in 15 TypeScript source files under `src/`; the Android wrapper adds 3 Java files.
 - The project is structurally split between framework-free gameplay modules in `src/game/` and Phaser scene code in `src/scenes/`.
 - `src/scenes/GameScene.ts` and `src/game/ai.ts` remain the two largest concentrations of application logic.
-- The AI transport layer is now a stable three-file path: `ai-worker-protocol.ts`, `ai-worker-client.ts`, and `ai.worker.ts`.
+- A dedicated audio preference path now exists: `src/game/preferences.ts`, `src/scenes/MenuScene.ts`, and `src/scenes/SettingsScene.ts`.
+- `main.ts` still contains a local `SettingsScene` placeholder class, while `MenuScene.ensureSettingsSceneAvailable()` swaps in the concrete imported scene before navigation.
+- The AI transport layer is a stable three-file path: `ai-worker-protocol.ts`, `ai-worker-client.ts`, and `ai.worker.ts`.
 - The AI exposes three difficulty levels: `beginner`, `advanced`, and `master`.
 - `advanced` and `master` both use `CardTracker` to reason about unseen cards without directly reading hidden hands.
-- The current `master` search profile is `timeBudgetMs: 4600`, `sampleCount: 10`, `maxDepth: 6`, `batchSize: 2`.
+- The base `master` search profile is `4300 ms / 8 samples / depth 5 / batch 2`, with tighter endgame branches down to `3200 ms / 4 samples / exact remaining depth / batch 1` when 4 cards remain.
 - `GameScene` consumes AI progress callbacks to update an on-screen think bar while a worker request is running.
+- `GameScene` now enforces `AI_MIN_THINK_MS = 1000` and `MOVE_OUTCOME_STATUS_MS = 2000` through timer-backed scene logic.
 - `AIWorkerClient` fails over pending work to in-thread `chooseMove()` if worker creation, posting, or deserialization fails.
+- The AI benchmark harness is now in source under `src/game/ai-benchmark.ts` and `src/game/ai-benchmark-fixtures.ts`, and `package.json` exposes it as `npm run benchmark:ai-quality`.
+- The current benchmark contract is iteration 5: 13 fixed fixtures, 6 critical concepts, and 48 self-play matches.
 - The Android wrapper targets SDK 36 with `minSdkVersion` 24 and applies immersive mode from the native activity.
-- Audio remains procedural via Web Audio; no dedicated audio asset pipeline is present in the source tree.
+- Audio remains procedural via Web Audio; there is still no dedicated audio asset pipeline in the source tree.
 - No ESLint or Prettier configuration is present.
 - The only repository-wide verification command supplied is `npx tsc --noEmit`.
 
 ## Potential Improvement Areas
 
-- `GameScene.ts` still centralizes layout, turn flow, HUD updates, effects, and audio in one scene class, which raises maintenance cost.
-- `ai.ts` still combines heuristic tiers, inference helpers, determinization, and alpha-beta evaluation in one module.
-- Worker transport is isolated cleanly, but progress rendering remains coupled to scene-level UI concerns.
-- A 4600 ms master search budget may still be noticeable on slower mobile devices even with batch yielding.
-- There is no dedicated automated rules or AI test suite beyond type-checking.
+- `GameScene.ts` still centralizes layout, turn flow, HUD updates, effects, audio, status messaging, and AI orchestration in one scene class.
+- `ai.ts` still combines heuristic tiers, inference helpers, determinization, move ordering, and alpha-beta evaluation in one module.
+- The current settings flow works, but the dual registration pattern for `SettingsScene` in `main.ts` plus dynamic replacement in `MenuScene` is fragile and worth simplifying later.
+- Worker transport is isolated cleanly, but progress rendering and fallback behavior remain coupled to scene-level UI concerns.
+- A 3.2 to 4.35 second master search window may still be noticeable on slower mobile devices even with yielding and the minimum-think pacing already in place.
+- There is no dedicated automated rules test suite beyond type-checking and the AI benchmark harness.
 - Formatting and style are enforced socially rather than by automated linting or formatting tools.
 
 ## Current Rule / Implementation Notes
@@ -39,14 +45,15 @@ Initializer refresh for SCOPONE-0009. The cached findings were stale relative to
 - When multiple direct matches exist, `findCaptures()` returns one single-card option per matching card.
 - Subset-sum captures are considered only when no direct match exists.
 - `applyMove()` defaults to the first legal capture if no explicit capture choice is supplied.
-- Scope is awarded only when a capture clears the table before the final play of the round.
+- Scopa is awarded only when a capture clears the table before the final play of the round.
 
 ### AI implementation snapshot
 
 - `beginner` uses a simpler heuristic with noise to remain beatable.
-- `advanced` adds race awareness, anti-scopa logic, partner setup, anchor play, and tracker-based probability estimates.
-- `master` orders legal moves with a quick evaluator, samples hidden hands, and scores them with alpha-beta search under the active deadline.
+- `advanced` adds race awareness, anti-scopa logic, partner setup, denari pressure, and tracker-based probability estimates.
+- `master` orders legal moves with a quick evaluator, samples hidden hands, and scores them with alpha-beta search under a dynamic deadline.
 - Progress is reported through `AIDecisionProgress` so the scene can keep the think bar responsive.
+- `CardTracker` now exposes same-rank residue summaries through `getValueRankResidue()` and `getValueRankResidueSummary()`, and those semantics are the live inference surface for unseen-value reasoning.
 
 ### Worker execution snapshot
 
@@ -59,11 +66,19 @@ Initializer refresh for SCOPONE-0009. The cached findings were stale relative to
 ### Scene / UI implementation snapshot
 
 - `BootScene` loads atlas assets and presents a simple loading bar.
-- `MenuScene` exposes difficulty selection before match start.
+- `MenuScene` now exposes both difficulty selection and a dedicated entry point into `SettingsScene`.
+- `SettingsScene` persists music and effects toggles immediately through `saveAudioPreferences()`.
+- `GameScene` reads normalized audio preferences from scene data or persisted storage before match start.
 - `GameScene` tracks played and captured cards in `CardTracker` as the round evolves.
-- The scene owns score HUD rendering, player labels, status text, think-bar rendering, and procedural particle effects.
+- The scene owns score HUD rendering, player labels, status text, think-bar rendering, procedural audio, and particle effects.
 - Round-end and match-end flows remain managed inside the scene instead of separate overlay components.
 
+### Benchmark snapshot
+
+- `ai-benchmark.ts` now uses a simulated timing source for fixture and self-play evaluation instead of depending only on wall-clock timing.
+- The benchmark summary records per-seed aggregates, dual-loss seeds, and a regression watchlist intersection.
+- The harness remains source-local under `src/`, so it is covered by the default `npx tsc --noEmit` include set.
+
 ## Research Performed
 
 ### Web Research: Scopone Scientifico Rules (2026-03-31)
@@ -118,34 +133,11 @@ Initializer refresh for SCOPONE-0009. The cached findings were stale relative to
 - The current `GameScene` pattern of registering one-shot shutdown and destroy handlers is aligned with Phaser guidance for worker disposal and UI cleanup.
 - Dealer rotation and next-round state changes can stay inside the existing in-scene orchestration without requiring a different Phaser lifecycle primitive.
 
-### SCOPONE-0009: Iteration 3 strength-planning notes (2026-04-08)
+### SCOPONE-0010: UI, settings, and benchmark refresh notes (2026-04-09)
 
-- `src/game/ai.ts` currently generates master determinization samples by uniformly shuffling all unseen cards and slicing them into opponents' hidden hands; it does not yet bias assignments by dealer role, parity residue, or observed capture semantics.
-- The transposition-table key in `src/game/ai.ts` includes the exact sampled hidden hands, so reuse is effective within a determinized sample but does not merge equivalent uncertainty classes across different sample assignments.
-- No executable benchmark harness or AI quality test module exists under `src/`; the current timing evidence lives only in prompt artifacts such as `prompts/SCOPONE-0009/iteration_2/benchmark_summary.md`.
-- `tsconfig.json` includes only `src`, so any automated quality or self-play harness that should be typechecked by the default `npx tsc --noEmit` command needs to live under `src/` unless the project configuration changes.
-
-### SCOPONE-0009: Iteration 3 continuation notes (2026-04-09)
-
-- The accepted iteration 3 benchmark work is now present in source: `src/game/ai-benchmark.ts` and `src/game/ai-benchmark-fixtures.ts` exist under `src/`, `package.json` exposes `benchmark:ai-quality`, and the harness already measures fixed fixtures, self-play, and production-master timing.
-- The live production master budgets in `src/game/ai.ts` are already below the requested five-second ceiling in every shipped branch: base `4300`, `<= 20 cards` `4350`, `<= 12 cards` `4200`, `<= 8 cards` `3900`, `<= 6 cards` `3600`, and `<= 4 cards` `3200` milliseconds.
-- `src/scenes/GameScene.ts` still executes AI turns immediately after `await aiClient.chooseMove(...)` resolves in `doAIMove()`; there is currently no scene-level minimum think-time floor.
-- `src/scenes/GameScene.ts` still uses a bare `setStatus(msg)` helper that only calls `this.statusText.setText(msg)`; there is no timed persistence policy, no cancellation of prior status timers, and no dedicated post-move outcome message path.
-- Phaser 3.87 scene timers can be cancelled with `TimerEvent.remove()` and their references cleaned with `TimerEvent.destroy()`; the current scene already listens to `shutdown` and `destroy`, so timed status cleanup belongs in the existing `handleSceneShutdown()` path.
-
-### SCOPONE-0009: Iteration 3 refresh notes (2026-04-09)
-
-- The current `src/game/ai.ts` heuristic does not reason about numeric even/odd card values; it already computes the unseen copy count for each rank and stores whether the remaining copies for that rank are in a singleton residue or a paired residue, but the internal names still use `oddResidue`, `evenResidue`, and `scoreParityTableState`, which can mislead future work.
-- The live tactical seam that needs refresh is therefore naming and policy framing, not a wholesale replacement of the underlying signal: the AI should explicitly treat `apparigliare` / `sparigliare` as preserving or breaking same-rank copy residues and connect that to table control, scopa prevention, and forced replies.
-- The accepted benchmark harness in `src/game/ai-benchmark.ts` still measures runtime with `performance.now()` and therefore depends on wall-clock search time. It does not yet use an injected or simulated search clock for fast validation runs.
-- `src/scenes/GameScene.ts` already contains the previously planned pacing and status work: `AI_MIN_THINK_MS = 1000`, `MOVE_OUTCOME_STATUS_MS = 2000`, a timer-backed `setStatus(...)`, and `handleSceneShutdown()` timer cleanup are all present in source and should be preserved rather than re-planned.
-- `src/game/ai-benchmark-fixtures.ts` still contains one fixture and tag using the stale label `dealer-parity-preserve-pair` / `critical-dealer-parity`; if benchmark files are reopened for simulated timing, that terminology should be refreshed to rank-residue wording at the same time.
-
-### SCOPONE-0009: Iteration 5 planning notes (2026-04-09)
-
-- The live AI quality harness in `src/game/ai-benchmark.ts` still hard-codes an `iteration: 4` quality gate with targets of `12` fixed fixtures, `4` critical concepts, and `48` self-play matches requiring `>= 30` wins and `<= 12` losses; the readable summary does not yet surface cross-seed aggregation such as the recurring dual-loss seeds from the latest rejected run.
-- `src/game/ai-benchmark-fixtures.ts` currently covers `settebello-capture`, `anti-scopa-defense`, `dealer-rank-residue-preservation`, and `exact-endgame-resolution` as critical concepts, but it does not yet encode an explicit critical fixture for partner invitation / partner scopa setup and does not yet make `fare scopa` itself a critical concept despite the user's new ordering.
-- Non-critical fixtures already exist for denari pressure, late denari shielding, and seven pressure, so the benchmark seam for iteration 5 is to rebalance critical-vs-fixed coverage and ordering expectations rather than to introduce a second harness.
-- Cross-tier heuristic priorities are concentrated in `src/game/ai.ts`: beginner logic in `scoreCaptureBeginner()` / `scoreDumpBeginner()`, advanced logic in `scoreCaptureAdv()` / `scoreDumpAdv()`, and master root/search logic in `quickEval()`, `orderSearchMoves()`, `generateSamples()`, and `evaluateFast()`.
-- Partner-aware logic already exists in all three tiers, but it is currently additive and distributed across multiple heuristics; there is no single explicit priority ladder that guarantees `partner setup` outranks seven denial, denari denial, and generic material capture across the whole file.
-- Anti-scopa prevention is already strong enough to pass the fixed tactical fixtures, but the rejected iteration 4 result (`18` wins, `30` losses over `48` seeded self-play matches) indicates that full-game strength is still limited by strategic continuity across seed-intrinsic lines rather than by isolated tactical blindness.
+- `src/game/preferences.ts` is now the authoritative audio preference seam. It normalizes stored values and shields scenes from malformed storage state.
+- `src/scenes/MenuScene.ts` now reads persisted audio preferences and exposes a dedicated settings entry point instead of keeping audio options implicit.
+- `src/scenes/SettingsScene.ts` exists as a real scene and persists music and effects toggles independently through `saveAudioPreferences()`.
+- `src/scenes/GameScene.ts` already contains the previously planned pacing and status work: `AI_MIN_THINK_MS = 1000`, `MOVE_OUTCOME_STATUS_MS = 2000`, timer-backed `setStatus(...)`, and `handleSceneShutdown()` timer cleanup are all present in source and should be treated as current behavior, not future work.
+- `src/game/ai-benchmark.ts` now enforces an iteration 5 contract with simulated timing, cross-seed aggregation, dual-loss reporting, and a regression watchlist intersection. Older findings that described iteration 4 targets or wall-clock-only timing are stale.
+- `main.ts` still registers a local `SettingsScene` stub while `MenuScene` dynamically installs the concrete scene implementation before use. This works today but is an architectural wrinkle worth remembering in later planning.