test(zippychord): property-based test suite (documents a backspace under-count bug)

[nightscape]

Summary

Adds a stateful property-based test suite for zippychord (using proptest-state-machine), plus two small bug fixes the testing surfaced. Treats zippychord as the state machine it is: a generated reference model drives a real Kanata SUT and asserts the net typed text after every transition.

Draft / WIP: the new property tests are RED on purpose — they document a real, pre-existing zippychord backspace under-count bug that is not fixed here. The tests are the spec the fix must satisfy; the fix is intended as a follow-up. CI will show these failing.

Commits

1. fix(zippychord) – zchd_reset didn't clear zchd_same_hold_activation_count (only cleared on a clean all-keys-released). A reset mid-activation left it stale, and it gates the common-prefix backspacing logic.
2. test(sim_tests) – clear the global PRESSED_KEYS at the start of the sim helper, so the new state-machine test (which drives that global and can panic mid-scenario while shrinking) can't contaminate other tests.
3. test(zippychord) – the property-based suite itself.

The oracle (why this isn't circular)

The oracle is split between generator and reference: a ChordExpansion transition carries which chord it activates (so the expansion text is known by construction), while the reference state only does coarse placement (fresh append / followup replace / disabled passthrough / free-typing literal). It deliberately does not reimplement the keystroke-level eager/overlap/backspace accounting — that's the code under test — so the bug surfaces instead of being mirrored.

The bug it documents (RED)

When an activation reuses characters from a prior eager activation via the common-prefix optimization, those reused chars aren't counted for deletion, so a following overlapping/followup activation backspaces too few characters and leaves stray text.

• Deterministic minimal repro (repro_overlap_underdelete): dict b→"c", b→"cfbcc", bd→"fee "; pressing b, SPACE, d yields "cfee " instead of "fee ".
• The state machine independently shrinks to a fresh 3-chord overlap; the committed proptest regression seed reproduces it deterministically.

Notes

• Coverage and the deferred dimensions (⌫-in-output, smart-space full punctuation, modifiers/caps-word, multi-key followups, …) are documented in src/tests/sim_tests/ZIPPY_PBT_NOTES.md.
• The PBT also caught the two issues fixed in commits 1–2, and (during development) demonstrated that a naive one-line fix to the under-count is insufficient — there are multiple manifestations.

🤖 Generated with Claude Code

https://claude.ai/code/session_01D2JYVbtxCmQMMDLL39FeHz

[nightscape]

Should we delete minimisation unit tests?
proptest-state-machine stores the seeds of failing tests, so these would reproduce as well.

[nightscape]

Should we delete minimisation unit tests?
proptest-state-machine stores the seeds of failing tests, so these would reproduce as well.

[nightscape]

Give this a longer readable name.

[nightscape]

Imo ChordExpansion should be even more granular: A Vector of (time_after_initial_keypress, KeyPress / KeyRelease)