fix(reliability): keep pause/cancel failures visible and recoverable

[Wangnov]

Summary

• Keep pause/cancel/discard failures visible inside the active progress surface as structured role="alert" feedback, with distinct copy for delivery failure, backend refusal, and an operation that has already entered an uninterruptible phase.
• Preserve the actual task state when stop requests fail: controls return to a retryable state, a paused discard failure remains paused with both resume and retry-cancel available, and late completions from an older request cannot overwrite a newer operation.
• Bind every pause/cancel IPC to the exact backend operation token and serialize ownership on macOS and Windows, so a delayed request from operation A cannot arm operation B. Every native/window quit path now atomically arms cancellation before commit; nested guards preserve that latch, updates check it after the phase boundary, and non-resumable uninstall/purge work blocks quit throughout.
• Add symmetric macOS/Windows regression coverage for false/rejected pause and cancel requests, failed paused-cache discard, stale completion events, and high-concurrency operation-lock behavior.

Validation

• npm run check — pass
• npm run lint — pass
• npm test — pass (20 files / 135 tests)
• npm run build — pass
• cargo clippy --manifest-path src-tauri/Cargo.toml --workspace --all-targets -- -D warnings — pass
• cargo test --manifest-path src-tauri/Cargo.toml --workspace — pass (97 tests)
• operation-lock stress regression, 100 iterations with --test-threads=64 — pass
• confirmed-quit/commit-boundary stress regression, 100 iterations — pass
• destructive-token claim vs. quit stress regression, 100 iterations — pass
• git diff --check origin/main...HEAD — pass
• Exact codex review --base origin/main — pass: no P1/P2 and no findings (未发现明确且可操作的功能性缺陷。)
• GitHub PR checks — pass (6/6)

Deployment notes

None. This changes runtime recovery behavior only and does not require a persisted-data migration.

Closes #168

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.