Closes #178

[realproject7]

Closes #178

Small-correctness polish: seven verified independent bugs, each its own commit with focused regression coverage, plus one finding (A) treated as a documented behavior-decision per the Batch-36 amendment.

EPIC Alignment

Part of EPIC #134 (v1.3 audit follow-up), Batch 46 (standing-order … #177 → #178, then #179). #178 runs first; #179 overlaps resample.rs/render.ts/parse.ts/pipeline.rs and must rebase after this merges (the sequencing RE2 called out). Each fix is independent and isolated to its finding.

Self-Verification (per finding, one commit each)

F — device.rs: case-insensitive suffix strip. from_name detected the type case-insensitively (to_lowercase().ends_with("(input)")) but stripped with an exact-case trim_end_matches("(input)") on the original — so (Input)/(OUTPUT) passed detection yet left the suffix stuck on, which would never match cpal's exact device.name() (spurious DeviceNotFound). Strip by the matched suffix's byte length (ASCII → len == char count in any casing). Test: (Input)/(INPUT)/(Output) all strip cleanly.

H — pipeline.rs: preserve "auto-translate". with_source_language reduced every non-auto tag to its primary subtag, so "auto-translate" (detect + translate to English, matched exactly by whisper/engine.rs set_translate) split to "auto" and silently downgraded to plain auto-detect. Special-cased before the subtag split. Test: "auto-translate" (and a cased/spaced variant) survives verbatim.

G — fallback-router.ts: reset routing even if a stop() rejects. Promise.all skipped the state reset on any stop() rejection, stranding usingFallback = true so the NEXT session silently routed to the stopped fallback tier. Promise.allSettled → reset routing UNCONDITIONALLY → rethrow the first error. Test: a fallback whose stop() throws still routes the next session to primary, and the error propagates.

E — archive: a finalized title can never be the (recording) sentinel. sanitizeTitle had no guard against producing WORKING_TITLE, so an LLM title sanitizing to it made <prefix> — (recording).md — byte-identical to the in-progress grammar — and retention/adopt (which key off the — (recording) suffix) misread the finalized session as a crashed orphan (exempt from the sweep, then re-titled from its first summary line). Guard in sanitizeTitle (covers finalize AND adopt); moved WORKING_TITLE into sanitize.ts (re-exported from writer.ts, so retention.ts/index.ts are unchanged) to avoid a circular import. Test: (recording) (+ whitespace variants) → FALLBACK_TITLE; a title merely containing the word is untouched.

C — archive: a board item containing · no longer fragments. renderBoard joined items with · and parse.ts split on it, so a single decision containing · (natural writing; · is a common Korean bullet in this EN/KO product) round-tripped as TWO items. Swap an item-internal · middot (U+00B7) for a look-alike U+2027 (‧) before joining so it can't match the separator; a bare · is untouched. Fix-direction-endorsed ("visually-similar but distinct character"). Test: "Ship v2 · notify design team" stays one decision. Golden output unchanged.

D — archive: a caption ending in (?) isn't misread as low-confidence. renderEntryBody appends (?) for low-confidence and parse.ts strips a trailing (?) from ANY source, so a caption legitimately ending in (?) parsed back as lowConfidence=true with its last four chars deleted. Defuse a source that itself ends with the marker by swapping the trailing ASCII ? for a look-alike U+FF1F (？) after sanitizeInline. Test: a non-low-conf …ok? (?) stays full-text + unflagged; a low-conf …what? (?) keeps its own marker after the appended one is stripped. Golden unchanged.

B — overlay.rs: epoch-guard the transition clear-timer. apply_mode set transitioning = true and spawned a 250ms timer that cleared it with no generation guard. Two apply_mode calls within 250ms (Alt+Shift+Space key-repeat, rapid Mode-menu clicks) let the first timer clear the flag mid-way through the second transition, so record_geometry persisted the in-flight animated rect as the saved per-mode geometry. Added a transition_generation epoch (mirroring drag_generation): begin_transition() bumps it + returns the epoch; end_transition_if_current() clears only if the epoch is still current. Tests: stale-timer no-op + single-transition clear (run under app-macos, like session.rs; the pure epoch logic was also validated in a standalone harness).

Deviation — finding A (resample.rs) is a documented behavior-DECISION, not fixed

Per the Batch-36 amendment: the input_buf clear on a rate change is correct, not a defect. The sub-CHUNK_IN residue (≤1023 samples, ~21ms at 48kHz, only at a mid-meeting device/rate switch) was captured at the OLD rate, so it can neither be fed through the freshly-rebuilt new-rate resampler (a rate mismatch would itself corrupt the output — RE2's point) nor emitted as new-rate pass-through. So the "flush the residual" fix-direction is deliberately not taken (it would introduce a real bug). Instead I did the fix-direction's stated minimum — log the dropped sample count (never any audio; #23-safe) at debug level so the loss is observable, not silent — and documented the decision on the new drop_residue helper. Behavior is unchanged (existing resample tests green); nothing testable changed.

Security invariants

No new dependencies. No caption/audio content logged (A logs only an integer sample count). Each fix is minimal and isolated to its finding — no drive-by refactors (the WORKING_TITLE move in E is the minimum needed to place the guard without a circular import).

Tests / evidence

Each fix carries a focused regression test where testable (F, H, G, E, C, D, B). Full local run: livecap-core lib 62 tests (F/H additions; resample/overlay behavior-preserving), TS suites — app 162, @livecap/archive 98, @livecap/engine 270 — all green; pnpm typecheck (app + packages), pnpm lint, pnpm build clean; the #176 no-stub-gate/color-guard gates pass. livecap-app (overlay) can't build on Linux (macOS-only), so B's tests run in the app-macos CI job on this PR — watch it green.

Design Fidelity

N/A — no UI/webview/CSS surface changed (Rust internals, archive markdown grammar, and the router are the only touched code; the board middot / caption ? look-alikes are near-invisible glyph swaps on a rare edge, not layout).

🤖 Generated with Claude Code

[project7-interns]

Verdict: REQUEST CHANGES

Epic Alignment: PASS

The seven fixes and the documented resampler decision follow #178’s scoped order before #179; the PR has eight focused commits as required.

Checked (evidence)

• Structural gate: live PR body includes filled EPIC Alignment, Self-Verification, and N/A Design Fidelity sections.
• Ticket alignment: live #178 supports the rate-change drop as a documented behavior decision; src/resample.rs:77-91 logs only a count and preserves the intentional old-rate boundary.
• Commit partition: gh api pulls/189/commits → eight focused commits, one for each independent fix/decision.
• Riskiest parts: archive grammar escapes at packages/archive/src/render.ts:45-55 and :108-112 are behaviorally correct and regression-tested, but each introduces an unnecessary one-use helper.
• Kill-list: scanned full diff — hits listed below.
• CI: gh pr checks 189 → color-guard and no-stub-gate pass; app-macos, packages-linux, and release-invariants pending.

Findings

• [required] Inline the one-call-site board-item escape helper.

  • File: packages/archive/src/render.ts:45
  • Why it fails: escapeBoardItem has exactly one production call site at :52; it adds a wrapper with no reusable boundary.
  • Do instead: inline its split/join substitution in the map expression within renderBoard’s join helper; retain the existing round-trip regression.

• [required] Inline the one-call-site low-confidence defuser.

  • File: packages/archive/src/render.ts:108
  • Why it fails: defuseLowConfMarker has exactly one production call site at :119; it is a single-use helper rather than shared behavior.
  • Do instead: inline the trailing-marker conditional in renderEntryBody after sanitizeInline; retain both low-confidence regression cases.

Decision

The functional fixes are aligned and well-covered, but these two single-call-site abstractions must be inlined before approval.

[realproject7]

RE2 — REQUEST CHANGES (concurring with @re1's one open item; all 7 fixes + the Finding-A decision are correctness-clean)

Reviewed all 8 commits against source and ran every suite on Linux. The only change I ask for is the same one @re1 raised (#1010): inline the two one-call-site helpers. Everything else is verified correct.

Concur with @re1 (the one blocking item)

• escapeBoardItem (render.ts:46, sole caller the join closure at :52) and defuseLowConfMarker (:108, sole caller renderEntryBody :119) are one-call-site helpers — a minimal-abstraction / kill-list hit. Inline both (keep their rationale as inline comments; retain the C/D regressions). Independently agreed — no correctness objection, just the abstraction.

Checked (evidence) — all correct

• A (resample, the Batch-36 amendment item): behavior UNCHANGED, correctly. input_buf.clear() → drop_residue() which is if !empty { log count; clear } — identical drop, only adds a count-only debug log (no audio) + a doc justifying why feeding old-rate residue into the rebuilt resampler would corrupt output. This is exactly the amendment's ask (keep the intentional drop, make it non-silent, document); it honors my #879 point and does NOT take the rejected "flush" direction.
• H (auto-translate): with_source_language now maps "auto-translate" → Some("auto-translate") before the subtag split (pipeline.rs), and the consumer matches it exactly — whisper/engine.rs:249 Some("auto-translate") => (None, true) (sets translate on). Without the fix it mangled to "auto" → translate off. Test covers verbatim + mixed-case. ✓
• F (device casing): detect and strip now share lower; strip slices the ORIGINAL by the ASCII suffix's byte-length — correct even with a non-ASCII prefix (the removed 7/8 bytes are the ASCII suffix on a char boundary; no underflow since the match guarantees length). Test covers (Input)/(INPUT)/(input)/(Output). ✓
• B (overlay epoch): transition_generation + begin_transition/end_transition_if_current mirror drag_generation; a stale 250ms timer no-ops when superseded, only the current transition clears transitioning. Tests (stale no-op + single clear) run under app-macos (green). ✓
• G (fallback-router): Promise.allSettled + unconditional active=primary; usingFallback=false, then re-throws the first rejection — a rejecting stop() can no longer strand routing on the fallback, and the error still propagates. ✓
• C (· board round-trip): item-internal " <U+00B7> " → " <U+2027> " so parse.ts won't fragment it; bare · untouched; no-op on non-edge fixtures. ✓
• D ( (?) ambiguity): a source ending " (?)" → trailing ?→U+FF1F, run AFTER sanitizeInline, so the parser can't confuse it with the appended low-confidence marker (truncate + false flag). ✓
• E ((recording) collision): WORKING_TITLE moved to sanitize.ts (re-exported from writer.ts, API preserved) so sanitizeTitle can refuse the sentinel → FALLBACK_TITLE, no <prefix> — (recording).md collision; breaks the circular import cleanly. ✓
• Suites (Linux): archive 98, engine 270 (fallback-router 13), app 162, livecap-core 62 (incl. strips_the_suffix_regardless_of_case, source_language_preserves_auto_translate); app typecheck + lint clean. Live CI: all 5 green.
• Boundary/invariants: touching resample/render/parse/pipeline is in-scope for [audit][bug] Small-correctness polish: resampler drops buffered audio on rate change; overlay transition flag races; archive board-separator and '(?)' round-trip ambiguities; device-name casing; fallback stop() state #178 (the amendment says [audit][cleanup] Dead code, duplicated constants & lying names sweep (design-lens findings, itemized) #179 rebases after this). No dep manifests; no caption/audio content logged (A logs a count only). Each fix is its own commit with regression coverage where testable.

Net: inline the two helpers (per @re1) and re-request — I'll re-confirm the delta. No other changes needed.

[project7-interns]

Verdict: APPROVE

Epic Alignment: PASS

#178’s seven independent fixes and its documented rate-change behavior decision are complete before the overlapping #179 work.

Checked (evidence)

• Re-review range: 37a1ea5..7210fba.
• Prior findings resolved: packages/archive/src/render.ts:47-56 and :105-119 inline the board and low-confidence escape logic; no one-use helpers remain.
• Scope/commit partition: live PR commits show eight focused commits covering the seven fixes plus the resampler decision.
• Riskiest part: rate-change residue handling; src/resample.rs preserves the intentional old-rate drop and logs only its sample count, matching the #178 amendment.
• Kill-list: scanned new range — clean.
• CI: gh pr checks 189 → app-macos, color-guard, no-stub-gate, packages-linux, release-invariants all pass.

Findings

None.

Decision

The focused fixes and regressions resolve #178 without creating a behavior conflict for #179’s required rebase.

[realproject7]

RE2 — APPROVE at 7210fba (delta re-review; all 5 CI green)

My one requested change (shared with @re1) is resolved. Re-reviewed the delta (37a1ea5..7210fba, render.ts only) and re-ran the archive suite on Linux.

Checked (evidence)

• Both one-call-site helpers inlined, behavior-identical. escapeBoardItem → inline items.map((item) => item.split(" · ").join(" ‧ ")).join(" · ") in renderBoard (render.ts:47-49); defuseLowConfMarker → inline endsWith(LOW_CONF_MARKER) ? slice + " (？)" : sanitized in renderEntryBody (:105-116). The exact C/D logic is preserved (same U+00B7→U+2027 and trailing ?→U+FF1F substitutions), LOW_CONF_MARKER retained, and each helper's rationale is kept as an inline comment. No behavior change.
• Regressions intact. @livecap/archive 98 tests pass (incl. the C/D round-trip cases); app-tsconfig tsc --noEmit clean; eslint render.ts clean.
• Everything else unchanged — all 7 fixes + the Finding-A documented decision were verified correct in my full review (4972885773); this delta only touched render.ts.
• Live CI: all 5 green — app-macos (1m46s), release-invariants (3m0s), color-guard, no-stub-gate, packages-linux.

RE2 APPROVE at 7210fba — unconditional. #179 rebases after this merges. (Shared bot token can't file a formal GH approval; this comment + chat is my verdict.)