A7: Unify backspace into one revertible input-unit stack

[AsafMah]

Part of #16. Depends on the A3 harness (most bug-prone area).

Today two backspace-pop mechanisms coexist: tryFragmentBackspace (manual spacing) and the combining whole-word delete. Model each input (tap/fragment/gesture-word) as a pushed input-unit with revert metadata; backspace pops one unit and restores composing/committed state. Migrate both onto the stack and delete the parallel paths (the area that produced the corruption PR #11 hand-fixed).

Acceptance: backspace predictable across manual/combining/default; corpus + unit tests green.

[AsafMah]

Closing as substantially done.

Shipped:

• test(backspace): golden-master backspace corpus for #31 (#21) #72 — JUnit backspace golden-master corpus (A3b: JUnit replay harness + golden corpus #21), the regression net for this area (composing-side: per-char, fragment-pop, combining whole-word, the "precool" recompose-corruption guard, monotonicity).
• refactor(inputlogic): consolidate backspace state into BackspaceUnitStack (#31) #77 — BackspaceUnitStack: the three scattered backspace state fields (mGestureFragmentBoundaries + mLastGestureCommittedLength + mLastGestureCommittedFragmentLengths) consolidated into one unit-tested object (18 direct unit tests on the pop/commit math). Behavior-preserving; verified on-device.

Deferred (the control-flow collapse): not worth it. Examined handleBackspaceEvent for a plan -> apply collapse of the deletion branches. They are not a clean dispatch — they're heterogeneous handlers interleaved with cancel-combining/stroke resets, the shortcut-expansion revert, shift-update timing, the recorrection commit, suggestion restart, and stats, each with different side effects and load-bearing early-returns. A plan type rich enough to encode them + an apply switch would mirror the branches ~1:1 — same complexity, added indirection, real regression risk in the most corruption-prone method, for marginal readability. Abstraction that doesn't pull its weight.

Future option (if ever revisited): a low-risk readability pass extracting the four deletion branches into named private methods (rejectBatchComposingWord / deleteWholeComposingWord / popCommittedGestureFragment / deleteCommittedGestureWord) with control flow unchanged. Not the original "one uniform pop" — that's incompatible with the heterogeneous handlers here.