Stability cleanup: remove falsified two-thumb experiment code from production paths #147

Why

main / v0.3.0 is stable, but dev currently carries ~1,950 added lines from the two-thumb experiment in #141. Device testing falsified the central production claim:

• DUAL_POINTER produces nonsense for real words such as "ambulance".
• The research harness measured the in-repo AOSP decoder, but the feature is visible only when a user-supplied/system glide library is loaded — a different, closed binary.
• IdealPrefixTrailBuilder was also actively harmful when no native library was loaded, until the UI was gated.

Known-bad default-off code is still complexity in the hot input path. Leaving it around indefinitely is exactly how an experimental fork becomes unmaintainable.

Keep — proven or independently useful

• PointerIdNormalizer and its tests: fixes a real zero-suggestions bug when no input point carries id 0.
• Side-by-side experimental build type (.exp) and distinct IME label: useful infrastructure for future device A/B tests.
• Historical findings in docs/issues, including the falsification and why the measured decoder was not the running decoder.

Remove from production paths

• DUAL_POINTER mode and its preference/UI.
• IdealPrefixTrailBuilder and its preference/UI.
• Re-timing/gap/interval preferences.
• StrokeAligner indirection if its only remaining purpose is those modes; restore the previous connector merge directly in WordComposer rather than preserving an abstraction around a dead experiment.
• Native two-pointer host harness assertions that imply relevance to the user-supplied decoder. Keep only if clearly labelled as testing the unused in-repo engine and it provides ongoing value.
• Settings/search wiring and strings for removed controls.

Definition of done

• Normal/default typing path is as close as practical to pre-Two-thumb experimental build (upstream v4.1.2 merged, installs side by side) #141 complexity.
• No known-bad two-track or synthetic-prefix options are exposed.
• Proven pointer-id normalization remains.
• .exp packaging remains for future controlled experiments.
• Existing regression corpus and native suite pass; full JVM results match the checked-in baseline through tools/check_test_results.py.
• Device smoke: normal swipe, two-thumb connector behavior, and no zero-suggestion case after pointer id 0 lifts.

Ordering

Land after the LeanBitLab v4.1.8 sync and the double-tap Shift fix, so the cleanup is based on the tree we intend to stabilize rather than creating another rebase/conflict round.