feat(testing): pointer-trace recorder + capture format (#20)

[AsafMah]

Implements #20 (A3a) — a debug capture mode that dumps real gesture sessions to replayable trace files, building toward the JUnit replay harness (#21) and, through it, the ability to verify the gated spacing/recognition epics.

What

When the new debug pref PREF_RECORD_INPUT_TRACES (Settings → Debug, default off) is on, each completed gesture/batch session is written to filesDir/input_traces/trace-<epochMillis>.json.

Format (v1, documented in TraceRecorder KDoc)

{ "version":1, "createdAt":<ms>,
  "keyboard": {"width":<px>,"height":<px>,"mainLayout":"qwerty","locale":"<bcp47>"},
  "committedWord":"<word>",
  "pointers":[{"id":<int>,"x":<int>,"y":<int>,"t":<ms>}, ...] }

pointers is the InputPointers flattened in index order — the exact trace fed to the recognizer (a merged trail for multi-part gestures), which is the most faithful replay input.

Implementation notes

• Hook: InputLogic.onUpdateTailBatchInputCompleted — chosen over onEndBatchInput because that's where the committed word (composedText), mWordComposer.getInputPointers(), and the keyboard are all synchronously available on the UI thread (the word is computed async by InputLogicHandler).
• Thread-safety: TraceRecorder copies the pointer arrays synchronously on the caller thread (InputPointers is mutable/reused), then writes off a single-threaded executor. Uses the application context (no IME-service leak).
• Cost: default off → one boolean check on the gesture-commit path; nothing in normal use.

Verification

• Compiles (Java+Kotlin); :app:testOfflineRunTestsUnitTest = 4 failed, all the Windows-only ParserTest set — zero new failures.
• On-device: flip the debug toggle, swipe some words, pull input_traces/*.json — not yet exercised end-to-end by me; worth a quick check when convenient.

Next: #21 (JUnit replay harness + golden corpus) consumes this format. Built by a subagent; coordinator reviewed (thread-safety, hook correctness), compiled, ran the suite. Base dev.