fix: enable Vim editing in question answers - #86

Self-review of 7101f9b

Verdict: fix the palette-ordering issue before merge. CI is green, but the following reproduction exposes a correctness bug outside the current test suite.

P1: answer keys overtake a queued command palette

src/index.ts:202-207

In an answer containing hello world, enter normal mode, then send :x followed by Enter in one terminal write. The palette callback is outside the answer's deferred-action queue, so x deletes the final d and Enter commits the edited draft before the palette opens.

This reproduces on both modeled host paths and on cache-installed OpenCode 2.0.15. Closing the palette reveals Answer: hello worl on the form confirmation screen. The form is still pending until a separate confirmation; this is an unintended draft edit/commit, not an immediate complete question submission.

Order the answer's pending palette action with its other actions, preserve the palette's exemption from focus-loss cancellation, and re-check host ownership before routing the following key. Add a coalesced palette/edit/Enter regression alongside the existing focus-change survival test.

P2: the undo-isolation test accepts a shared stack

test/question-answer.test.ts:134-154

The test deletes in main, deletes in the answer, undoes the answer, then undoes main. A shared LIFO stack returns the same snapshots in that order. Redirecting all snapshot clear/push/pop operations to the main context in a disposable copy still leaves all 75 answer tests passing.

Undo main first while the answer still has its own snapshot, then refocus and undo the same answer object. Assert both buffers throughout.

P2: the toggle-reset test replaces the editor

test/question-answer.test.ts:392-405

Calling open() after the toggles creates a fresh editor, so the test cannot detect whether the old answer context was reset. Removing the context-reset line in a second disposable copy also leaves all 75 tests passing.

Keep the same answer editor focused across disable/re-enable, seed normal mode or a pending operator, and assert that subsequent typing uses fresh insert state.

The P2 findings are test-oracle gaps, not current production bugs. No additional blocking security, packaging, clean-code or style issues were found. The small shared context module and unchanged pure engine are appropriate for this scope.

Validation remains 430 passing tests, 16 harness-unit checks, and 39 v1 / 56 v2 installed-host checks. Live v1 question editing is still a declared coverage limit. The inherited tab-offset/snapshot-redo gaps and existing typecheck errors are outside this review. No fixes were applied during this review.

Review findings addressed in c36d9f2

All three self-review findings are fixed. The final independent re-review found no remaining issues, and CI passes on the updated PR.

• Palette transitions preserve input order across asynchronous mounting and focus changes. Both keys and bracketed paste wait for palette focus, then pass through the native host input APIs. Buffered input cannot edit or commit the previous answer after ownership is lost.
• Undo-isolation tests now use non-LIFO editor switches and an answer with no local snapshot. They reject a deliberately shared undo stack.
• Toggle-reset tests retain the same answer editor with normal mode, a pending operator/count and an undo snapshot. They reject removal of the context reset.

Final validation:

• just check: lint clean, 448 tests pass, including 93 shared answer cases.
• just compat-unit: 16 tests pass.
• Fresh cold/warm cache installs of the same artifact: 39 checks pass on OpenCode 1.18.33; 62 pass on 2.0.15. Native v2 palette-key and bracketed-paste bursts preserve the exact answer and produce the expected palette query.
• Mutation checks reject both previously undetected test regressions.

The two inherited tab-offset/snapshot-redo gaps and six existing typecheck errors remain unchanged. Live v1 question-field coverage remains the declared validation limit.

Ready for human review. No merge or release performed.