fix: cancel queued messages and preserve question answers

[andrebrait]

Stack: #67 → #88

Stacked on #67. Merge #67 first; this PR adds queue cancellation and question-answer preservation on top of its native queue-promotion implementation.

• Base layer: fix: promote queued follow-ups through native RPC #67 — native queued-message promotion.
• This layer: fix: cancel queued messages and preserve question answers #88 — acknowledged queue Delete/Edit and question-dialog fixes.

Both PRs currently target main, so this PR’s Files changed tab also includes the #67 implementation. After #67 merges, rebase this branch onto main so only this layer remains in the diff.

GitHub’s native stack registration is unavailable for these cross-fork PRs: it currently requires all branches to be in the same repository. This section records the explicit dependency and merge order; it does not indicate native stack membership.

Summary

• Make queue Delete/Edit acknowledge native cancellation before removing the chip or recalling text. Preserve the original draft across navigation/unmount and publish one queue snapshot to remounted same-session hooks without double-removing duplicates.
• Preserve question-dialog text and selections when SSE reconnects replay the same request ID; reset only for a genuinely new question.
• Update localized confirmation text and refusal handling; failed cancellation leaves the queued item visible.

Contributor's reported symptoms:

Deleting messages I steered doesn't seem to be actually deleting them. They get eventually delivered as follow-ups instead.

The contributor also reported intermittent clearing of the question dialog's free-text answer. The cause was a reset effect keyed to request-object identity rather than question ID.

Explicit prerequisites and overlap

Draft: native RPC prerequisites are not yet in an official OMP release. Cancellation is provided by can1357/oh-my-pi#11872. Promotion is provided by can1357/oh-my-pi#11618.

This PR builds on the existing implementation proposed in #67: real native promotion, ref-serialized queue updates, delivery-before-acknowledgement bookkeeping and the StrictMode alive-state correction. It is the next layer, not a replacement for #67. Native Delete cannot safely follow a cosmetic-only Steer label because that would address the wrong backend queue. Review and merge #67 first, then the additional changes here; this does not introduce unrelated layout, provider, selection or integration patches.

Unsupported runtimes report an error and keep the queue item. There is no abort/resend fallback.

Verification

• Upstream source port: typecheck, lint and full suite passed: 711 passed, 1 existing skip.
• Regression coverage includes acknowledged/failed cancellation, duplicate queue isolation, same-session remount synchronization, original-draft recovery across navigation/unmount, promotion races and real question replay/new-ID behavior.
• Original dialog replay regressions erased text/selection against baseline and passed after the fix. Original remount regressions retained ghost chips against baseline and passed after acknowledgement publication.
• Agent-operated packaged-app/native runtime test reset actual staging TCP connections, observed SSE reconnect, and submitted the exact preserved multiline answer to native OMP. The contributor also exercised the deployed question field.
• Compiled integrated stack test: follow-up -> native Steer -> Delete removed the target before it reached the provider; a remaining follow-up was delivered normally. The standalone native upstream port has its own real-RPC smoke.
• The combined downstream integration is deployed locally with old active chats preserved during runtime retirement. The deployed integration includes other patches; this upstream diff does not.
• Single condensed adversarial review of the upstream port found no blocking code defects.

Copilot/CodeRabbit review requests are routed only to companion PRs in our forks, not to this upstream PR.

Companion automated-review PR (Copilot/CodeRabbit run only in our fork): andrebrait#21

Promotion acknowledgement correction

Successful native promotions now update the session-scoped persisted mirror even after unmount/navigation. Shared pending ownership also protects a remounted delivery-before-ACK case from promoting the next duplicate. The remount, observer, subsequent Delete, duplicate, and unobserved-session regressions failed before this correction and pass afterward. Both feature variants pass 41 focused hook/composer/dialog tests; the refreshed clean deployment passed 682 tests with one existing skip. All README feature descriptions now match native cancellation semantics.

Final review corrections

• Same-key remounted composers receive recalled text without losing pending typing or duplicating it.
• Queue cancellation/promotion ACK waits are bounded to 5 seconds, with no automatic retry or process termination.
• A vanished local queue item reports an actionable warning.
• A separate failed-submit path was also found: the old finally timer dismissed a question after a rejected response. It now dismisses only after success, preserving the pending question and answer for retry. The original periodic while-typing issue was same-ID SSE replay, not this timer.

Verification: 58 focused hook/composer/dialog/draft/real-RpcProcess tests pass, plus TypeScript and ESLint, in both the upstream port and integration feature. The test-renderer migration nit remains with existing upstream #85 rather than adding a second test framework in this fix.

[noor-latif]

I noticed this issue too! Impressive PR. I'm telling you, you could join the OMP dev team. Would have made this PR smaller :)

[andrebrait]

@noor-latif I do have a few PRs open to bring the required changes to OMP too :)

[andrebrait]

@noor-latif and part of the reason this PR looks so big is because it is technically stacked on top of #67. Let me make that explicit :)