feat(spacing): signal-driven grace duration + live knobs (#24)

[AsafMah]

Phase 2 increment of the spacing-policy epic (#14) — implements #24's signal-driven grace + the first #26 live knobs.

What

The combining grace-timer duration now adapts to the #92 word-state signals instead of a fixed value:

graceMs = clamp(base − completeBonus·complete + prefixPenalty·prefixRichScore, 100, 3000)

• a finished dictionary word (complete) commits sooner (− completeBonus);
• an extendable prefix-rich stem waits longer (+ prefixPenalty·score).

Wired into enterCombiningMode behind the default-off flag PREF_SPACING_SIGNAL_DRIVEN_GRACE. The formula is a pure static helper (signalDrivenGraceMs) for unit-testing; base = the existing PREF_COMBINING_GRACE_MS; the min/max clamp is hardcoded (100–3000 ms).

Live knobs (#26) — experimental, on the Two-thumb screen

• Adapt pause to the word (toggle, default off)
• Finished-word speed-up = completeBonus (slider, default 200 ms)
• Extendable-stem patience = prefixPenalty (slider, default 400 ms)

Verification

SpacingSignalsTest (+6 signalDrivenGraceMs cases: neutral=base, complete shortens, prefix-rich lengthens, combined, floor/ceiling clamp) / SettingsContainerTest / InputLogicTest → 128 completed, 3 failed (the AGENTS.md pre-existing baseline), 1 skipped — 0 new failures. Default-off → existing grace behavior unchanged.

⏳ On-device: enable "Adapt pause to the word" (with the grace timer on), type, and tune the two sliders by feel — finished words should commit a touch faster, extendable stems linger. Building the .dbgtest APK.

Next: the two-gate Assisted tier (Gate A instant / Gate B pause) + the A11 insight readout.

[AsafMah]

Parking/closing after on-device testing changed the product conclusion.

The adaptive timer lever is weaker than expected: tap-only stems are now protected by Only auto-finish swiped words, while pure swipes usually resolve to a full word (not a prefix stem), so the prefix-richness patience path is hard to exercise/feel. This adds knobs and behavior complexity without touching the main problem often enough.

Keep the underlying signal layer from #92; revisit only if we find a concrete surface where adaptive timer changes are observable and valuable (likely multipart/tap-then-swipe), not as a broad user-facing feature.