[dzianisv]

Fixes #197

Ports the speech hunks from open PR #182 standalone — a mechanical port of already-reviewed hunks, not a fresh design. None of #182's other fixes are included, and #182's speech-errors.test.ts is not ported.

Diff summary

• New src/lib/speech-errors.ts — pure, import-free helper isBenignSpeechError(code) returning true for 'no-speech' and 'aborted'. Single place classifying benign speech errors. Its doc comment states only what is documented: abort() emits aborted, and events are module-global.
• src/lib/speech.ts — error handler uses isBenignSpeechError(event.error) instead of the inline 'no-speech' check, so a module-global aborted from another screen's cleanup no longer sets an error; unmount cleanup only calls abort() if this hook instance actually started listening, tracked with a listeningRef synced via useEffect (not assignment during render); the misleading abort() is a no-op when not listening comment is corrected in place.

Why

Backing out of a subagent session showed a phantom "Voice input failed" alert on the parent screen with the mic never used: the child screen's unmount cleanup called abort(), the module-global aborted event reached the parent's still-mounted error listener, and only no-speech was carved out. Real voice errors still surface — only no-speech and aborted are treated as benign.

Gate (local checkout, branch head a8b376e, base e844493)

• npm ci — added 794 packages in 45s
• npm run check:versions — Version metadata aligned: 0.4.17 (44)
• npm run typecheck — clean (tsc --noEmit, exit 0)
• npm test — 380 pass, 0 fail

Context: issue thread comment 6855175743 (context pack) and design-gap comment 6855230505 on #197. No release/tag from this PR — release is a separate act after merges land.

[dzianisv]

Stage 4 review — PASS

PR: #235 — fix(speech): treat aborted as benign; only abort when listening (#197)
Head SHA: a8b376e41d3ffab70d3d02b2a8e1acc43bb415df (verified by API readback and local checkout)
Test-count delta: 0 — test files 42 → 42 (no test files added or removed)

How reviewed

Read the full #197 thread, context pack (comment 6855175743), design-gap comment (6855230505), and the PR diff, then reviewed the real checkout at the head SHA. Note: ~/workspace/custom-skills/claude-code-review/SKILL.md does not exist on this VM (no matching skill found), so the review followed the task method directly: acceptance criteria checked one by one against the diff, each evidence claim independently validated.

Acceptance criteria — all verified

1. No phantom alert on back-nav without voice use — two layers: unmount cleanup now skips abort() when this hook instance never started listening, and even if an aborted event arrives, it is classified benign and never reaches setError. (On-device confirmation remains a post-merge manual residual, as the pack notes — no e2e harness for the native speech module exists.)
2. Real voice errors still alert — the else branch is unchanged: setError(event.message || event.error), and the alert consumer in app/session/[id].tsx is untouched.
3. aborted benign, like no-speech — isBenignSpeechError returns true for exactly "no-speech" and "aborted" (src/lib/speech-errors.ts).
4. Cleanup aborts only when this hook instance actually started listening — listeningRef is synced via useEffect on [listening], never written during render; the unmount cleanup reads listeningRef.current.
5. No changes outside useSpeech / speech error handling — diff is exactly 2 files, +28/−4: new src/lib/speech-errors.ts, src/lib/speech.ts edits. No applicationId change, no other files touched.

Checks that passed

• Misleading comment corrected — the false "abort() is a no-op when not listening" is replaced with the corrected wording.
• Helper doc comment limited to documented claims — no unverified Kotlin "ALWAYS emits" assertion. The "aborted is the documented outcome of calling abort()" claim checks out against the library README: "You'll also receive an error event (with code "aborted") when calling .abort()" and error-code table "aborted — The user called ExpoSpeechRecognitionModule.abort()".
• Gates re-run on the PR head by this reviewer (not the implementer's numbers): npm run typecheck clean (exit 0); check:versions aligned 0.4.17 (44); full suite npm test — 380 pass / 0 fail.

Findings

No findings. No validated defect, so nothing blocks merge.

Notes (not findings)

• The speech.ts cleanup comment asserts the Kotlin-level detail that abort() emits the global event even when idle — Stage 3 already flagged that as reported-but-unverified. It does not affect correctness (the benign classification covers the event either way), so it is not a finding.
• CI at observation time: Typecheck and unit tests ✅, test ✅, Secret scan (gitleaks) ✅; build, Unsigned iOS Simulator build, and activation-e2e still in progress; build-fdroid skipped. Combined status: pending.