Rewrite SessionController's concurrency model onto coroutines/StateFlow #8

Objective

Replace SessionController's five hand-coordinated Handler threads (bg, capture, speech, plus main, plus the accessibility service's own thread) and manual @Volatile/turn-id state with structured concurrency (Kotlin coroutines + StateFlow/Mutex), without changing the frozen OverlayCommand/planner contracts or any of the five core pieces.

Rationale

Flagged by the repo audit as the highest-maintainability-risk code in the project. The turn-invalidation logic (newTurn(), isCurrent(), recordingTurn) is correct today because it was carefully hand-verified this session, but every new feature added to this class (app-choice state, chosenPackage, currentChoices, the newer open_ended/AppLauncher path) grows the same unstructured state space. This is not a bug report - nothing is currently broken - it's a bet against future regressions being introduced silently.

Acceptance criteria

• No Handler/HandlerThread remains in SessionController - replaced by coroutines on appropriately-scoped dispatchers
• Turn invalidation (a mic-tap during "Thinking..." must not let a stale plan repaint the pill) is expressed via Job cancellation, not manual turn-id comparison
• Every existing behavior verified in this session's manual device testing still holds: stop-mid-recording, stop-and-resume preserving step position, cursor/speech running in parallel with bounds resolution, debug panel not clobbering itself
• All existing unit tests pass unmodified or are updated to test the same behavior through the new concurrency model
• No change to contracts/*.schema.json, OverlayCommand, or PlannerResult

Implementation steps

1. Introduce a viewModelScope-equivalent CoroutineScope owned by SessionController, cancelled in dispose()
2. Model a "turn" as a cancellable coroutine Job rather than an AtomicInteger comparison
3. Move WavRecorder start/stop behind a Mutex instead of a dedicated Handler thread
4. Move TTS synthesis + playback onto its own dispatcher via withContext(Dispatchers.IO), preserving the "speech never blocks the visual path" property
5. Re-run every manual verification scenario from this session's transcript (mic-tap-during-thinking, Stop mid-utterance, taxi app-choice flow) on a physical device before merging

Estimated effort

High (3-5 days for one engineer familiar with the codebase, including manual device re-verification of every documented scenario)

Risks

• This is the single riskiest change achievable in this codebase - it touches the class every other feature depends on
• Coroutine cancellation semantics differ subtly from manual turn-id checks; a rushed port could reintroduce the exact stale-turn bug this session fixed (PR Make a voice turn stale-safe and actually observable #2)
• No instrumented tests exist today to catch a regression automatically - this PR should add some as part of the work, not just port the logic

Testing checklist

• Unit tests for turn cancellation semantics
• Manual device test: mic-tap, then mic-tap again during "Thinking..." - pill must stay on the second turn, first turn's plan must never render
• Manual device test: Stop mid-recording, then mic-tap again - must resume from the preserved step, not restart
• Manual device test: full taxi flow (choose app -> launch -> guide inside it) end to end
• Regression: keyless deterministic fallback path still works with no Sarvam key configured