B4: Per-thumb pointer attribution (true simultaneous two-thumb)

[AsafMah]

Part of #15. Gate on the A3 harness.

When one thumb glides and the other taps simultaneously, all touch points are aggregated into one stroke and the tap warps the recognized word. Keep the two thumbs' point streams separate by pointer ID (infra exists: PointerTrackerQueue): gliding thumb -> stroke; tapping thumb -> discrete taps via live-converge.

Acceptance: a tap mid-glide no longer corrupts the swiped word in corpus tests.

[AsafMah]

This is the actual lever, and it is probably mis-classified as separate from #97.

PR #136 found that the AOSP decoder already models two pointer tracks (MAX_POINTER_COUNT_G == 2) and can spell a word alternating between them. We have never used the second one: InputPointers.appendAll hardcodes pointer id 0 (its own comment notes multi-part composition doesn't preserve pointer identity). That hardcoded 0 is why the connector hack exists at all.

Because updateTouchPoints partitions purely on pointer id, per-thumb attribution is the composition mechanism — not a different problem. #97 currently classifies this as "a different problem, not composition"; that separation looks wrong.

#136 also fixes a real bug found here: if no point carries id 0, Suggest::initializeSearch early-returns and you get zero suggestions — reachable in ordinary two-thumb use (A down, B down, A lifts, B swipes on as id 1).