[Epic] Composing word = editor text as single source of truth (two-thumb refactor)

[SHAWNERZZ]

Tracking epic for the multi-phase refactor of two-thumb / multi-part composition. Design: docs/COMPOSING_WORD_SOURCE_OF_TRUTH.md.

Goal: make the editor text + composing region the single source of truth for the current word, and derive stroke geometry from text on demand — removing the parallel, drift-prone state (mLiveStroke, the mInputPointers-survives-reset contract, mGestureFragmentBoundaries) that causes the recurring stale-stroke / desync bugs. Unlocks mid-word cursor editing and partial deletes.

Phases (each independently shippable + on-device validated):

• Phase 1 — Stop the bleeding in fragment mode. On a fragment-pop, realign mInputPointers to the truncated word so a following swipe-extend doesn't merge with the pre-pop stroke. (Closes the fragment case the fresh-word reset in Two-thumb: gesture stroke isn't cleared on backspace — re-typing a mis-recognized word accumulates geometry (words grow each retry) #7/fix(two-thumb): clear stale stroke buffer so re-typed words don't accumulate geometry (#7) #9 doesn't cover.)
• Phase 2 — Prove the model. Make live-converge derive its merge base from the current word's text (getCoordinatesForCurrentKeyboard) instead of mLiveStroke; validate re-recognition on-device; retire mLiveStroke.
• Phase 3 — Generalize. Route all multi-part composition through 'composing region as source of truth': derive fragment boundaries from text, support mid-word cursor + partial-delete continuation, retire mGestureFragmentBoundaries.

Related: builds on #7/#9 (fresh-word stroke reset). Primary risk: synthesized key-center strokes vs real captured strokes — validate recognition on-device at each phase.

[SHAWNERZZ]

Scope clarified in the design doc (PR #10) per review — explicit Trigger model:

• Taps are exact and bypass live-composition entirely (this replaces the current re-recognize-taps feature, PREF_MULTIPART_RERECOGNIZE_TAPS — taps no longer fold into the stroke).
• Moving the cursor into a word is context-only — it must not live-compose / re-recognize / alter the word.
• Only a swipe (re)composes, building onto whatever word is at the cursor (typed, swiped, or just moved into) via its text-derived key-center stroke.

Phase 2 now also drops tap re-recognition / mLiveStroke; Phase 3 adds 'swipe onto a word the cursor was moved into' with the trigger model strictly enforced.

[SHAWNERZZ]

Correction to my earlier scope comment (PR #10 updated): re-recognition is the core feature and is preserved — I previously over-corrected toward removing tap re-recognition, which was wrong.

Corrected Trigger model:

• A swipe always re-recognizes, pulling in the word at the cursor as context (typed / swiped / just-moved-into). Doc end + swipe ument → Document.
• A tap is exact UNLESS the current word already contains a swipe — then taps before/during/after contribute (the combine-taps-and-swipes feature stays). Do|c + tap g → Dogc.
• Cursor move is context-only (inert until a swipe; a bare tap stays exact).

mLiveStroke is retired only as stored state — the re-recognition behavior is preserved, re-sourced from the current word's text on demand.