mokosh/.planning/STATE.md

---
gsd_state_version: 1.0
milestone: v2.0.0
milestone_name: milestone
status: executing
stopped_at: Plan 01-12 closed via Wave 7 operator brand-fit ack 2026-05-20 'all good' (147/147 vitest, 21/21 UAT GREEN); Plan 01-10 (welcome tab) remains as final Phase 1 functional plan
last_updated: "2026-05-20T08:00:00.000Z"
last_activity: 2026-05-20
progress:
  total_phases: 5
  completed_phases: 0
  total_plans: 14
  completed_plans: 13
  percent: 93
---

# Project State

## Project Reference

See: .planning/PROJECT.md (updated 2026-05-15)

**Core value:** When an operator hits a bug, one click MUST produce a
self-contained archive that lets support reproduce what happened — in under 5 s,
no server, no password leaks.
**Current focus:** Phase 1 — Stabilize Video Pipeline

## Current Position

Phase: 1 of 5 (Stabilize Video Pipeline) — Plan 01-12 closed-by-brand-fit-ack 2026-05-20; Plan 01-13 closed-by-harness 2026-05-19; awaiting Plan 01-10 (welcome tab) as final Phase 1 functional plan
Next phase: 2 of 5 (Stabilize DOM + event-capture privacy)
Plan: 13 of 13 complete (13 plans: 01-01..01-09 + 01-11 spike + 01-12 + 01-13 + 01-14; Plan 01-10 still pending)
Status: Executing — Plan 01-10 (welcome tab) remains
Last activity: 2026-05-20

Progress: [█████████▓] 93%

### Plan 01-12 closure (2026-05-20)

- Design integration landed end-to-end across 7 waves (10 plan tasks + 1 operator empirical checkpoint)
- 9 implementation commits: `3fe018b` (plan baseline revision post-01-14) → `34a9ce1` (Wave 0 RED scaffolds) → `f86fd60` + `abab6e1` (Wave 1 fonts + tokens.css) → `7732a30` (Wave 2 icons) → `110cebc` (Wave 3 manifest i18n + _locales) → `468f16d` (Wave 4 source adoption) → `e8d2881` (Wave 5 Vite define + welcome conditional) → `b909c37` (Wave 6 A18-A22) + 1 pre-checkpoint commit `865d394`
- SUMMARY: `f319c7d` (`.planning/phases/01-stabilize-video-pipeline/01-12-SUMMARY.md`)
- **R2 designer substitution**: Newsreader → Lora (Cyreal foundry; OFL-1.1; full Latin + Cyrillic) per designer reply 2026-05-19. Canonical token value `--mks-font-display: "Lora", "Iowan Old Style", "Times New Roman", serif`.
- **MV3 CSP self-host invariant verified**: zero `googleapis` / `https://fonts` references in `dist/` (per `tests/build/no-remote-fonts.test.ts`); 8 local @font-face rules in `src/shared/tokens.css`; ~155 KB font bundle ships under `src/shared/fonts/` with LICENSE + README attribution.
- **16 i18n keys across `_locales/{en,ru}/messages.json`**: extName + extDesc + tooltipOff + tooltipRecPrefix + tooltipErr + popupSavePrompt + popupSaveCta + popupSaveDone + popupSaving + popupSaveDoneShort + popupEmptyState + popupInfoText + notifStartup + notifRecovery + welcomeHeroRu + welcomeHeroEn. EN extName = "Mokosh — Session Capture" (D-07); EN extDesc = "Thirty seconds ago, always at hand." (D-08); RU extName = "Mokosh — Запись сессии"; RU extDesc = "Тридцать секунд назад, всегда под рукой." en↔ru parity verified by `tests/i18n/locale-parity.test.ts`.
- **Branded Loom-mark icons** (D-01): 8-bit RGBA via `rsvg-convert`; replaces Bug A placeholders (was 16-bit RGB at 574/1153/2615 B; now 406/784/1952 B 8-bit RGBA). Clears Chrome imageUtil silent-rejection floors (16≥200B, 48≥500B, 128≥1024B).
- **BADGE_REC_COLOR** flipped from material-green `#00C853` to madder `#b2543d` (= --mks-madder-600 per D-04 loom palette).
- **src/popup/style.css** carries ZERO hex literals (every color via `var(--mks-*)`); imports `../shared/tokens.css`. src/popup/index.ts + src/background/index.ts read `chrome.i18n.getMessage('<key>') || '<en-const-fallback>'` at every operator-facing site.
- **UAT harness A18-A22** following Plan 01-13 Approach B pattern (page-side `assertA*` + host-side `driveA*` + harness orchestrator). FORBIDDEN_HOOK_STRINGS at 13 entries (+1 over 01-14's 12 baseline for `data-mks-key` completeness). Full A1-A14 + A18-A22 + A23 chain runs in ~95s under Puppeteer headless.
- **Pre-checkpoint bundle gates** per feedback-pre-checkpoint-bundle-gates.md established: SW CSP grep (new Function/eval) + SW Node-globals grep (Buffer.*) + DOM-globals grep + manifest validation + en↔ru parity. Discovery: setimmediate polyfill `new Function` in SW chunk via `vite-plugin-node-polyfills` — VERIFIED pre-existing across Phase 1 history; NOT a Plan 01-12 regression; logged at `.planning/phases/01-stabilize-video-pipeline/deferred-items.md` for Phase 5 hardening.
- **vitest: 100 → 147 GREEN** (+47 across 6 new test files at tests/build/ + tests/i18n/).
- **UAT harness: 16/16 → 21/21 GREEN** (A18-A22 added; A22 skip-gates on Plan 01-10 absent).
- **Operator empirical brand-fit ack 2026-05-20 verbatim "all good"** on fresh build + load unpacked + branded-surface verification (toolbar Loom icon, popup loom palette + Lora display heading, manifest:name resolution to "Mokosh — Session Capture", Russian copy rendering with Lora, notification copy via chrome.i18n).
- **Plan 01-13 Task 9 (operator brand/design ack on loaded extension) functionally CLOSED** via this checkpoint — same operator + same empirical surface coverage; the LAST remaining Phase 1 brand-design gate.

### Outstanding Phase 1 gates

- **Plan 01-10 (welcome tab):** still pending — operator-facing onboarding surface (Russian copy; first-run only). Canonical `src/shared/tokens.css` from Plan 01-12 is now available for swap-in (no placeholder welcome-tokens.css needed). When Plan 01-10 lands, harness A22 flips from skip-gate to 1-check GREEN.

### Plan 01-13 closure (2026-05-19; brand/design ack subsequently closed via Plan 01-12 Wave 7 2026-05-20)

- Puppeteer-based UAT harness: `npm run test:uat` exits 0 with **15/15 GREEN** (A0-A14)
- Bug A regression rewind empirically verified (commit body 6a77967)
- Bug B regression rewind empirically verified (commit body b665919)
- Plan 01-09 functional contract closed via harness PASS per `01-09-PLAN.md` Amendment 2
- Operator UAT Task 9 ack'd 2026-05-19 ("all good" — recovery + restart-after-click covered by harness A7 + A2)
- **Save-stops-recording charter divergence fixed inline via debug session** (`.planning/debug/resolved/01-09-save-stops-recording.md`):
  - Symptom: SAVE created zip but did NOT stop recording (badge stayed REC; Chrome share banner persisted)
  - Root cause: implementation 01-09 over-extended "always-on safety net" framing; SPEC intent is one-shot
  - Fix: SW SAVE_ARCHIVE handler dispatches STOP_RECORDING + setIdleMode in finally (4f4c3e2)
  - Harness regression coverage: A14 added (2b6c24b) — post-SAVE state check (badge='', popup='', no new recovery notif)
- **CHARTER REVERSAL 2026-05-19 — save-does-not-stop-recording** (`.planning/debug/resolved/01-09-save-does-not-stop-recording.md`):
  - Operator UX iteration: prefers original "always-on safety net" framing (continuous recording; SAVE only creates a new zip)
  - Revert: SW SAVE_ARCHIVE `finally` block REMOVED (commit 7645765)
  - Test file inversions: `tests/background/save-archive-does-not-stop-recording.test.ts` (renamed via `git mv`, history preserved; commit 6ac23fd)
  - Harness A14 inverted to assert continuous-recording post-SAVE: badge='REC', popup endsWith popup.html, no new recovery notif (commit 1baaf45)
  - Plan 01-09 Amendment 3 landed documenting the reversed charter
  - vitest preserved at 98 GREEN; `npm run test:uat` preserved at 15/15 GREEN under inverted contract
- Plan 01-11 closed as spike-pivot (ba5474c SUMMARY); architecture lessons (no `await import(...)` in SW; `track.dispatchEvent('ended')` not `track.stop()`; `__MOKOSH_UAT__` Vite define-token) carried forward into Plan 01-13's Approach B harness
- vitest: 83 → 98 GREEN across Plan 01-13 (+15: Tier-1 grep gate strings + hook contract tests + save-stops unit tests)

### Outstanding Phase 1 gates

- ~~**Plan 01-13 Task 9 (operator checkpoint):**~~ CLOSED 2026-05-20 via Plan 01-12 Wave 7 brand-fit ack "all good"
- ~~**Plan 01-12 (design integration):**~~ CLOSED 2026-05-20 (R2 Lora substitution + tokens.css canonical + 16 i18n keys + branded icons + manifest i18n; SUMMARY f319c7d)
- **Plan 01-10 (welcome tab):** still pending — operator-facing onboarding surface (Russian copy; first-run only). Canonical `src/shared/tokens.css` from Plan 01-12 is now available for swap-in.

## Performance Metrics

**Velocity:**

- Total plans completed: 0
- Average duration: —
- Total execution time: —

**By Phase:**

| Phase | Plans | Total | Avg/Plan |
|-------|-------|-------|----------|
| 1. Stabilize video pipeline | 7 | ~50 min (+ 2 debug sessions ~45 min) | 7 min |
| 2. Stabilize DOM + event capture privacy | 0 | — | — |
| 3. Stabilize export pipeline | 0 | — | — |
| 4. SPEC §10 smoke verification | 0 | — | — |
| 5. Harden + clean up | 0 | — | — |

**Recent Trend:**

- Last 5 plans: 4min, 4min, 8min, 3min, ~10min (Plan 07 closure incl. debug-session arbitration)
- Trend: stable execution time; complexity surfaced in debug sessions (pre-staged fallbacks activated cleanly)

*Updated after each plan completion*
| Phase 01 P01 | 4min | 6 tasks | 6 files |
| Phase 01 P02 | 4min | 5 tasks | 8 files |
| Phase 1 P03 | 8min | 3 tasks | 5 files |
| Phase 01 P04 | 4min | 3 tasks | 1 files |
| Phase 01 P05 | 8min | 2 tasks | 1 files |
| Phase 1 P06 | 3min | 2 tasks | 2 files |
| Phase 1 P07 | ~10min closure + 2 debug sessions (D-12 + A3) | 2 tasks (checkpoint + auto) | 6 files (fixture + REQUIREMENTS + ROADMAP + STATE + SUMMARY + plan-final-commit) |
| Phase 01 P14 | 49m | 1 tasks | 7 files |
| Phase 01 P12 | ~10h cumulative (7 waves; 10 plan tasks + 1 operator empirical checkpoint) | 10 tasks (7 waves + Wave 7 pre-checkpoint + brand-fit ack) | ~50+ files (8 WOFF2 + 3 PNG + 2 _locales + tokens.css + 6 unit-test files + harness + scripts + 4 source files modified) |

## Accumulated Context

### Decisions

Decisions are logged in PROJECT.md Key Decisions table (DEC-001 through
DEC-012, all SPEC-Accepted and locked for Phase 1). Recent decisions affecting
current work:

- Phase 1 framing: roadmap treats the existing codebase as a partially-broken
  first attempt to be remediated against the SPEC, not as greenfield. The
  7 P0 defects from the audit are split across phases 1–3 along commit
  boundaries; phase 4 is end-to-end SPEC §10 smoke verification.

- All 12 SPEC decisions (`DEC-001`..`DEC-012`) are LOCKED for Phase 1.
  Changing any of them requires a formal ADR; none are formally LOCKED in the
  ingest classification, so a future ADR can revise.

- [Phase ?]: Doc cascade: amendments append (do not replace) original DEC/CON blocks to preserve SPEC provenance — Established convention for future SPEC-amending phases; downstream readers see both old + new with citation
- [Phase ?]: Manifest: drop alarms permission entirely rather than retain for re-use — Plan 05 deletes the alarms code path; declaring unused permissions expands attack surface (T-1-02)
- [Phase ?]: Pinned vitest at ^4 (4.1.6 latest stable; 5.x still beta on 2026-05-15)
- [Phase ?]: Phase 1 Wave-0 test infra: 4 RED tests committed against not-yet-existent src/offscreen/recorder.ts — pins contracts for Plans 03+04
- [Phase ?]: Reverted premature REQ-video-ring-buffer Complete marking left by Plan 01-01; satisfied by Plans 03+04+07, not by Wave-0 RED tests
- [Phase 01-03]: Bundled OffscreenLogger into Task 2 commit (Rule 3 blocking dependency — recorder.ts cannot typecheck without the import)
- [Phase 01-03]: Defensive bootstrap guard (typeof chrome check) lets pure ring-buffer test import recorder module without chrome stub
- [Phase 01-03]: Removed SW-side VIDEO_CHUNK/VIDEO_CHUNK_SAVED branches + IndexedDB helpers inline (tsc-clean requires; Plan 05 owns remaining SW shrink)
- [Phase 01-04]: Kept Plan 03's defensive bootstrap guard (typeof chrome / per-API existence checks) instead of Plan 04's verbatim unguarded block — Plan 04's verbatim block regressed ring-buffer and codec-check tests (they don't stub full chrome surface); restored guard preserves Plan 02 RED contract while satisfying Plan 04's new GREEN contract. Rule 1 deviation.
- [Phase 01-04]: T-1-04 SW-side sender check documented redundantly (4 places in recorder.ts) for Plan 05 executor visibility — Offscreen is trusting party; SW is validating party. Documenting in module header, port-name constant, threat-mitigation comment near bootstrap, and inline at connectPort makes the contract impossible to miss when grepping for T-1-04 during Plan 05.
- [Phase 01-04]: REFACTOR pass NOT skipped: stale 'Plan 04 wires this' comments replaced with actual D-17/Pattern 5 citations — Forward-pointing TODO-style comments became misleading after the work landed; minimal correctness-preserving comment update with all 9 tests still GREEN.
- [Phase ?]: [Phase 01-05]: Deleted broken checkPermissions / requestPermissions flow (Rule 1)
- [Phase ?]: [Phase 01-05]: REQUEST_PERMISSIONS collapsed — under getDisplayMedia (D-01) no runtime perm check is meaningful; the broken 'tabCapture' permission check was sending recording-start into the never-granted branch
- [Phase ?]: [Phase 01-05]: Added chrome.offscreen.hasDocument() in initialize() — Rule 2 robustness, audit P1 #8 mitigation across SW respawns
- [Phase ?]: [Phase 01-05]: SW is now a pure coordinator — onConnect host bound to 'video-keepalive' port with T-1-04 sender check; getVideoBufferFromOffscreen replaces synchronous SW-local buffer fetch; OFFSCREEN_READY handshake closes the audit P1 #12 race
- [Phase ?]: [Phase 01-05]: indexedDB.deleteDatabase('VideoRecorderDB') in onInstalled — T-1-NEW-05-02 / RESEARCH.md Runtime State Inventory cleanup of orphaned IDB from pre-Phase-01 builds
- [Phase ?]: [Phase 01-06]: Collapsed vite.config.ts from 226 -> 21 lines (RESEARCH.md Example B verbatim); deleted 174-line inline copy-offscreen plugin (audit P0 #1 root cause) and the orphan offscreen/ top-level directory (D-08)
- [Phase ?]: [Phase 01-06]: crxjs Outcome A confirmed — dist/src/offscreen/index.html (preserves src/ prefix from rollupOptions.input key). SW URL adjusted to chrome.runtime.getURL('src/offscreen/index.html'); RESEARCH.md Pitfall 5 binding empirically verified
- [Phase 01-07-debug-d12]: D-12 port-blob serialization fixed via base64 wire-format encode/decode (debug session d12-blob-port-transfer-fails resolved 2026-05-15). chrome.runtime.Port JSON-serializes payloads across extension contexts so Blob payloads were silently corrupted (JSON.stringify(blob) === "{}" → SW saw [{}, {}, ...] → new Blob([...]) coerced each to "[object Object]" → 75-byte text instead of WebM). Added src/shared/binary.ts (blobToBase64 / base64ToBlob), TransferredVideoChunk wire-format type, offscreen encode side, SW decode side. All 15 tests green incl. 6-test port-serialization spec. Re-run smoke.sh + ffprobe still required for end-to-end verification.
- [Phase 01-07-debug-a3]: D-13 restart-segments activated (debug session webm-playback-freeze resolved 2026-05-15). Plan 07 smoke retest after D-12 landed revealed the next-layer A3 failure: the ffprobe-valid WebM froze ~1 s into playback in Chrome because the single-continuous-recorder + 30 s age-trim lifecycle (D-09..D-11) evicted middle chunks containing VP9 keyframe references for retained tail chunks (orphan P-frames). Activated the pre-staged D-13 skeleton in src/offscreen/recorder.ts: stop+restart MediaRecorder every SEGMENT_DURATION_MS=10_000 ms on the same MediaStream, keep last MAX_SEGMENTS=3 self-contained WebM segments (3×10s=30s window preserved). Each segment fresh-encoded → own EBML header + seed keyframe → independently decodable. Side-effect: .stop() per segment fixes the "File ended prematurely" Matroska finalization gap. Type renames propagated: TransferredVideoChunk → TransferredVideoSegment, VideoChunk → VideoSegment, PortMessage.chunks → PortMessage.segments, VideoBufferResponse.chunks → VideoBufferResponse.segments; the header-pin flag from D-09..D-11 is dropped entirely. D-09..D-11 retired in favor of D-13. 28/30 tests pass; the 2 remaining reds are the empirical ffmpeg dry-runs against the still-stale committed fixture (operator regen required). REQ-video-ring-buffer NOT marked complete — Plan 07 still owns that, gated on the operator running ./smoke.sh then verifying Chrome playback + ffmpeg-clean stderr.
- [Phase 01-07-closure]: Phase 1 closed 2026-05-15: D-12 + A3 acceptance gates both passed. Operator-confirmed Chrome playback clean (no ~1 s freeze); ffmpeg `-v warning -i tests/fixtures/last_30sec.webm -f null -` exit 0 with zero decoder errors (only expected muxer DTS-monotonicity warnings at segment join boundaries — non-blocking, documented D-13 trade-off for multi-EBML-header concat); ffprobe + empirical playback both green; 30/30 vitest green (the 2 webm-playback empirical dry-runs flipped GREEN after the fresh fixture committed in cd61cbc); REQ-video-ring-buffer marked Complete; SPEC §10 #2, #3, #7 functionally satisfied (end-to-end Phase 4 smoke still owns the full §10 sweep). Three atomic closure commits land the fixture + REQ/STATE/ROADMAP flip + SUMMARY. Process note: Plan 01-07 surfaced TWO unanticipated-cascade failures (D-12 then A3); both had pre-staged fallbacks (base64 wire-format and D-13 restart-segments) that activated cleanly. Candidate retro: should `/gsd-plan-phase` auto-inject empirical-acceptance gates (ffmpeg dry-run + Chrome playback) before merging a phase when RESEARCH.md flags HIGH-risk assumptions?
- [Phase 01-07-deferred-to-5]: getDisplayMedia cursor visibility constraint (`video: { cursor: 'always' }`) surfaced as a user observation during Phase 1 smoke 2026-05-15. Captured frames lack the screen cursor despite it being the highest-signal cue for reproducing pointer-driven bugs. Constraint is opt-in per the getDisplayMedia spec; Chrome implements CursorCaptureConstraint (always/motion/never). Logged to Phase 5 P1/P2 hardening list — not blocking Phase 1 closure.
- [Phase ?]: Plan 01-14 — monitorTypeSurfaces:'include' shipped as top-level DisplayMediaStreamOptions constraint (W3C spec §6.1; Chrome ≥ 119 picker narrowing); A23 harness gate + Tier-1 grep lockstep extension to 12 strings; 100/100 vitest + 16/16 UAT GREEN. types.ts NOT modified — new cell/op are module-internal.
- [Phase 01-12]: Design integration landed end-to-end via 7 waves + operator brand-fit ack 2026-05-20 "all good". R2 designer substitution (Newsreader → Lora; Cyreal foundry; OFL-1.1; full Cyrillic via reply 2026-05-19) baked in; src/shared/tokens.css canonical with 8 local @font-face rules + zero remote URLs (MV3 CSP self-host invariant); 16 i18n keys per locale across en + ru with parity; branded Loom-mark icons replace Bug A placeholders (8-bit RGBA); src/popup + src/background migrated to chrome.i18n.getMessage with `|| <const>` fallback; BADGE_REC_COLOR flipped from material-green #00C853 to madder #b2543d (= --mks-madder-600 per D-04). UAT harness A18-A22 GREEN. Pre-checkpoint bundle gates established per feedback-pre-checkpoint-bundle-gates.md (5 grep gates pre-checkpoint; setimmediate polyfill new Function in SW chunk verified pre-existing across Phase 1 history — logged to deferred-items.md for Phase 5 hardening). vitest 100 → 147 GREEN (+47); UAT 16 → 21 GREEN (+A18-A22; A22 skip-gates on Plan 01-10 absent).
- [Phase 01-12]: Plan 01-13 Task 9 (operator brand/design ack on loaded extension) functionally closed via Plan 01-12 Wave 7 brand-fit ack 2026-05-20 (same operator + same empirical surface coverage). This was the LAST remaining Phase 1 brand-design gate.

### Pending Todos

None yet.

### Blockers/Concerns

- (informational) `chrome.tabCapture` requires a user gesture on first
  activation — Phase 3 (P0-4) restores this by moving the call into the popup
  click handler; until Phase 3 lands, recording cannot start cleanly even if
  Phase 1's pipeline is correct. Phases 1–3 should not be re-ordered.

## Deferred Items

Items acknowledged and carried forward from previous milestone close:

| Category | Item | Status | Deferred At |
|----------|------|--------|-------------|
| *(none)* | | | |

## Session Continuity

Last session: 2026-05-20T08:00:00.000Z
Stopped at: Plan 01-12 closed via Wave 7 operator brand-fit ack 2026-05-20 'all good' (SUMMARY f319c7d; 147/147 vitest + 21/21 UAT GREEN); Plan 01-10 (welcome tab) remains as final Phase 1 functional plan
Resume file: None

Prior session: 2026-05-19T19:41:05.737Z — Completed Plan 01-14 (commit b467123 + SUMMARY 5254145; 16/16 UAT + 100/100 vitest GREEN)
Earlier session: 2026-05-17T14:30:13Z — resumed from /gsd-pause-work checkpoint ed82fd6; Bug A icons (a881bf0) + intel-unlock (f768498) committed; /gsd-debug spawned for Bug B state-machine routing (subsequently resolved via the recovery-flow amendment at Plan 01-09 Task 5 step 11)

## Phase 1 Closure Notes

- **ffprobe exit code:** 0 (`ffprobe -v error -f matroska -i tests/fixtures/last_30sec.webm`)
- **ffmpeg dry-run exit code:** 0 (`ffmpeg -v warning -i tests/fixtures/last_30sec.webm -f null -`) — stderr contains only the expected muxer DTS-monotonicity warnings at segment join boundaries; no decoder errors. Documented D-13 trade-off for multi-EBML-header WebM concatenation; Chrome's MSE pipeline handles this natively (SPEC §10 #7 scope: "plays back in a browser" — Chrome confirmed).
- **Fixture:** `tests/fixtures/last_30sec.webm` = 1 633 459 bytes (1.6 MB), VP9 codec, Profile 0, 1142×1038, color space bt709, time_base 1/1000, start_pts 0. Captured against the D-13 restart-segments recorder (3 × ~10 s self-contained segments).
- **Test suite:** 30/30 green across 8 files (`tests/offscreen/`); both empirical ffmpeg dry-runs in `webm-playback.test.ts` flipped GREEN after the fresh fixture committed in cd61cbc.
- **Phase 1 outcome:** SPEC §10 acceptance criteria #2 (continuous capture), #3 (≤ 30 s window), and #7 (last_30sec.webm plays in a browser) are functionally green at the Phase 1 level. End-to-end §10 smoke verification remains owned by Phase 4 (all 9 criteria sweep).
- **Phase 2 onwards:** Phase 2 owns the DOM/event-capture privacy slice (REQ-rrweb-dom-buffer, REQ-user-event-log, REQ-password-confidentiality). Phase 3 owns the popup state machine + base64-URL replacement. Phase 4 runs the full SPEC §10 smoke pass. Phase 5 absorbs P1/P2 hardening (now includes the `getDisplayMedia` cursor visibility refinement surfaced 2026-05-15).
- **Process retro candidate:** Plan 07 surfaced two cascade failures (D-12 binary transfer + A3 cluster alignment). Both had pre-staged fallbacks (base64 wire-format and D-13 restart-segments) which activated cleanly. The smoke-test step ended up doing the empirical-acceptance-gate work that RESEARCH.md flagged as HIGH-risk. Worth raising in a GSD-framework retro: should `/gsd-plan-phase` auto-inject empirical-acceptance gates (ffmpeg dry-run + Chrome playback) BEFORE merging a phase when RESEARCH.md flags HIGH-risk assumptions, rather than discovering it via Plan 07's smoke step?