HomenShum · HomenShum · Jul 29, 2026 · Jul 28, 2026 · Jul 28, 2026 · Jul 28, 2026
diff --git a/.qa/HANDOFF-2026-07-28-walkthroughs-motion-proof.md b/.qa/HANDOFF-2026-07-28-walkthroughs-motion-proof.md
@@ -0,0 +1,135 @@
+# Handoff — 2026-07-28 · walkthroughs, CDP repair, motion-proof
+
+Written for the next engineer. Everything below was measured, not assumed; where something
+is unverified it says so. Three repos are involved, so read the **Landmines** section before
+touching anything.
+
+---
+
+## What shipped
+
+| Repo | PR | State |
+| --- | --- | --- |
+| FeatureClipStudio | [#6](https://github.com/HomenShum/FeatureClipStudio/pull/6) | **merged** 2026-07-29 |
+| NodeSlide | [#110](https://github.com/HomenShum/NodeSlide/pull/110) | mergeable, waiting on one CI check |
+| NodeRoom | [#244](https://github.com/HomenShum/NodeRoom/pull/244) | rebased onto main, CI re-running |
+
+### Three product defects, all found by filming the product rather than reading it
+
+1. **Every `FocusTrapDialog` modal rendered behind its own blur scrim.** Radix portals
+   overlay and content as *siblings*; the legacy scrims (`.r-modal-backdrop`,
+   `.r-room-modal-scrim`) were written to centre a *child*, and `FocusTrapDialog` passes
+   `unstyled`, which drops the Tailwind fixed/translate classes. Result: dialogs rendered
+   `position: static` at the end of `<body>`, under a z-95 blur — including the create-room
+   dialog carrying the product's one governance question. Fixed in `src/app/styles.css` by
+   positioning `[data-slot="dialog-content"]` itself, plus a real `.sr-only` (Tailwind's is
+   absent from this bundle, so the fallback `DialogTitle` printed the word "Dialog" on every
+   modal).
+   **DOM text extraction read these dialogs correctly the whole time.** Only frame review
+   caught it.
+
+2. **CDP scripts were killing the developer's real browser.** `browser.close()` on a
+   `connectOverCDP` connection closes *Chrome*, not the socket — so every script killed the
+   browser the next one needed. Swept the class: 9 pure-CDP scripts drop the call;
+   `scripts/motion-inventory.mjs` connects *or* falls back to `launch()`, so it tracks
+   `weLaunchedIt` and closes only what it owns. Separately `scripts/chrome-cdp-up.ps1` used
+   `Stop-Process -Force`, which is what produced *"Chrome didn't shut down correctly"* and lost
+   tabs; it now closes gracefully with `CloseMainWindow()`.
+
+3. **The upload verifier asserted videos that no longer exist.** `yt-verify.mjs` hardcoded two
+   ids that are now Private, and matched titles on a phrase the superseded clip *and its
+   replacement* both carry — so it could pass against the wrong video. Root cause was two
+   copies of the roster free to disagree. Now one `scripts/yt-roster.mjs` that both the
+   verifier and the privatize guard import, and the verifier checks **both directions**:
+   6 published must resolve, 4 superseded must be refused.
+
+### `skills/motion-proof` — vendored into this repo on purpose
+
+It previously lived only in `~/.claude`, which is **not version controlled**, so it could not
+be handed to anyone. It now sits beside `skills/liveflow`, `skills/probe-first`, etc.
+
+```bash
+node skills/motion-proof/motion-probe.mjs                    # the deception corpus
+node skills/motion-proof/motion-probe.mjs <url> --subject "<css>" --nudge
+```
+
+An audit found this skill **inverted in practice**: its `SKILL.md` correctly names
+`Element.getAnimations()` primary and the video judge advisory, but `getAnimations()` had
+**zero executable callers** while six Gemini video-judge scripts shipped. `motion-probe.mjs`
+is the missing primary instrument. It ships with 7 adversarial fixtures + an honest control;
+the control must pass and every deception must be caught, and running it the first time
+found two false positives in the probe itself.
+
+---
+
+## Landmines
+
+- **`~/.claude` is not a git repo.** `skills/motion-proof` is now vendored here, but the
+  *other* skills there (`design-dna`, `graph-hop`, `second-brain`, `trust-surfaces`,
+  `motion-ladder`) exist on one machine with no history and no backup. If they matter,
+  they need a home.
+- **PR #244 was stacked on `codex/nodekit-contract-alignment`.** That branch's own PR (#241)
+  fails `NodeSlide packed consumer`, `node-platform / conformance`, and `verify` — the same
+  three failures #244 inherited. It has been rebased onto `main` and those failures left with
+  it. **Do not re-branch from `codex/*` without checking its PR is green first.**
+- **Two conflicting PRs remain open in NodeRoom** — #190 (`codex/mobile-terracotta-launch`)
+  and #182 (`codex/proofloop-strict-live-official`), both `CONFLICTING`. Not touched; not mine.
+- **`.qa/memory/findings.jsonl` had an unresolved stash conflict** (`UU`). Both sides were
+  distinct valid records in an append-only log, so the resolution kept the **union** — 29
+  records, all parsing. If you expected one side to win, check that.
+- **Never render video live in a demo.** Image generation returns in seconds; video takes
+  minutes and fails often.
+- **Deferred boot.** `boot.ts` defers the app module until first interaction, so an
+  unhydrated SSR shell is a *different page* from the React landing — different markup, and
+  the join-code control is inline rather than a dialog. Any probe or capture must nudge and
+  wait for a React-only element, never a timer. This is why `motion-probe` has `--nudge`.
+
+---
+
+## The pattern worth inheriting
+
+The same defect appeared five times today in unrelated places: **a hand-typed value describing
+a version of the artifact that no longer exists.**
+
+- a YouTube title reading "11s walkthrough" over a 24-second video
+- `METADATA.md` recording pre-recut durations
+- showcase GIFs rendered from superseded captures
+- the verifier's hardcoded video ids
+- the roster existing twice, free to drift
+
+Every fix was the same shape: **derive the value at write time, or keep exactly one copy.**
+`yt-upload.mjs` now derives durations via `ffprobe`; `yt-roster.mjs` is the single roster.
+
+A second, sharper version of it: **a decision recorded only in a conversation will be
+re-decided by the code.** A council verdict to delete NodeSlide's Design tab as a standalone
+destination is not done, is recorded nowhere in that repo, and the most recent inspector
+commit reinforced the tab strip instead. Verdicts that survive must land in the repo they
+govern — as a test, an invariant, or at minimum a dated note.
+
+---
+
+## Verification habits used here (and why)
+
+- **Verify the claim the artifact makes, not that the artifact exists.** Frame counts, byte
+  sizes and HTTP 200s all passed while the clips showed another product's URL.
+- **Probe gates in both directions.** A gate only ever seen passing is not known to work. The
+  privatize guard was tested by feeding it a keeper and confirming it refused, *before* it was
+  aimed at real targets.
+- **Verify from the public surface.** `git push` exiting 0 is not evidence; fetch the raw URL
+  and grep for a content signal.
+- **State coverage, don't imply it.** Artifacts carry `touched/total`, and `JOURNEYS.md`
+  records 10 of 13 journeys shot with the remaining three named, one of them explicitly
+  declined by the owner rather than missed.
+
+---
+
+## Open, not done
+
+- NodeSlide #110 and NodeRoom #244 need a human to confirm the merge once CI is green.
+  NodeSlide's repo does not allow auto-merge, and `--admin` would bypass a branch-protection
+  rule that was set deliberately — so it was not used.
+- `ScoreReceipt` (third link of `ReferenceObservation → DesignRule → ScoreReceipt`) was being
+  emitted by a parallel agent when this was written; confirm it landed and that its schema
+  validates the existing records unmodified.
+- The 7 deception fixtures are runnable but are not yet wired into CI. They should be — that
+  is what turns the corpus from a demonstration into a gate.
diff --git a/.qa/memory/findings.jsonl b/.qa/memory/findings.jsonl
@@ -25,3 +25,5 @@
 {"severity":"P2","area":"mounted deck accessibility","symptom":"Two mounted workbench textareas had no accessible name","rootCause":"Placeholder text was used without an explicit label","fix":"Added aria-labels for slide comments and NodeAgent revision requests","status":"fixed","evidence":"docs/demo/nodeslide-i7-mounted-browser.receipt.json accessibility.unnamedInteractive=[]","ts":"2026-07-21T03:06:47.394Z","fp":"985afc2c40be"}
 {"id":"smb-lending-tablet-tour-overlay","sev":"P1","area":"responsive-layout","symptom":"tablet-first-run-tour-obscured-primary-workspace","rootCause":"desktop-onboarding-reused-at-tablet-breakpoint","evidence":"docs/release/proof/20260721-smb-lending-room/smb-lending-tablet-1024x768.png","fix":"binder-behind-explicit-Room-toggle","status":"fixed","ts":"2026-07-21T08:07:51.402Z","fp":"da997e793aa8"}
 {"id":"smb-lending-fixture-copy-drift","sev":"P1","area":"content-integrity","symptom":"agent-copy-disagreed-with-canonical-bank-statement-requirement","rootCause":"narrative-not-derived-from-fixture-label","evidence":"tests/smbLendingRoomSeed.test.ts","fix":"canonical-fixture-wording-and-seed-assertion","status":"fixed","ts":"2026-07-21T08:07:51.464Z","fp":"7c36fce153d0"}
+{"severity":"P0","area":"live agent write path (post-215)","symptom":"After #215 deployed, a fresh host-owned room spreadsheet task STILL writes nothing: GLM-5.2 calls write_locked_cells on a blank sheet (nothing locked) -> output-denied -> tool_blocked; no proposal created, review queue empty. #215 correctly capped the loop (stops step 8 needs-attention vs 20+ before) but the write still fails.","rootCause":"model tool-selection error: GLM-5.2 reaches for write_locked_cells instead of update_sheet/write_cells on unlocked cells. Denial has no recovery instruction routing it to the correct tool. This is model-quality-driven, confirming the Kimi K3 flip is the actual fix not just a preference.","fix":"PRIMARY: flip default to Kimi K3 (PR #216 + convex env AGENT_ORCHESTRATOR_MODEL/AGENT_MODEL). SECONDARY (follow-up): write_locked_cells denial on unlocked cells should return a recovery instruction naming the correct write tool, and/or the tool should no-op-fallthrough to update_sheet.","status":"open","evidence":"room NRKBC9322LN; progress step 4 tool-write_locked_cells output-denied; model nebius/zai-org/GLM-5.2; qa-memory 20d943680a7e refined","ts":"2026-07-18T10:31:37.668Z","fp":"395400d4c0bf"}
+{"severity":"P0","area":"live agent write path — TRUE root cause","symptom":"Fresh-room spreadsheet creation writes NOTHING, identically under BOTH GLM-5.2 and Kimi-K3. write_locked_cells returns tool_blocked failureKind=evidence_required stage=preflight_required.","rootCause":"verified_workbook_workflow guard (src/nodeagent/guardrails/workbookWorkflow.ts guardWrite) blocks any managed write until an inspect_workbook -> verify_workbook(afterWrite=false, full op set) -> write handshake establishes an approved plan. goalRequiresVerifiedWorkbookWorkflow() classifies ANY create/fill sheet-with-rows/columns goal as needing this heavyweight preflight — INCLUDING blank-sheet creation that has no existing data to protect. Neither model completes the handshake reliably (they write first, get preflight_required, do not cleanly recover), so blank-sheet creation is permanently blocked. This is the layer BENEATH what #215 fixed (A1-twin target coverage) — same deadlock family, different gate.","fix":"OPTION A (recommended): exempt genuinely blank/new sheets from the verified-workbook preflight (the workflow protects EDITS to existing/uploaded workbook data; fresh creation has nothing to verify against). OPTION B: server-side auto-run the inspect+verify preflight when a model calls write on a blank sheet (transparent recovery). OPTION C: strengthen system-prompt/tool-desc so models complete the handshake. Open question the user must confirm: is this preflight the SAME as the review-every-change wedge, or orthogonal (benchmark write-integrity vs human proposal approval)? If orthogonal, Option A is safe.","status":"open","evidence":"rooms NRKBC9322LN (GLM), NR8SN8GB3CR + NR0R5ZYG87A (kimi-k3); guardWrite preflight_required block payload captured; goalRequiresVerifiedWorkbookWorkflow matches the Q3-variance goal via rows/columns/create tokens","ts":"2026-07-18T10:43:44.401Z","fp":"84ea123117e2"}
diff --git a/AGENTS.md b/AGENTS.md
@@ -28,9 +28,13 @@ Use these files as the map:
 - `src/nodeagent/skills/integration/omnigentAdapter.ts` - Omnigent YAML compatibility checks.
 - `docs/NODEAGENT_ADOPTION.md` - porting checklist.
 - `docs/OMNIGENT_INTEGRATION.md` - Omnigent boundary and smoke command.
+- `skills/probe-first/SKILL.md` - refuse to draft outreach to a person without a verified, dated, source-backed research hook.
 
 Rules:
 
+- No outreach draft to a named person without a verified research hook on the
+  contact record. Enforce it on the tool-call path, not in a prompt; a prompt
+  rule loses under pressure. See `skills/probe-first/`.
 - Keep writes behind `RoomTools`; do not mutate engine/backend state directly in
   harness examples.
 - Keep durable memory in frames/cache/job rows, not prompt transcripts or

diff --git a/README.md b/README.md
@@ -843,10 +843,26 @@ result, with step captions and a progress bar. Regenerate and judge any time wit
 directly with `npm run walkthrough-review -- <feature-id> --ui-review`; lower-level
 capture/render commands remain `npm run walkthroughs` + `npm run walkthroughs:render`.
 
+### The NodeRoom walkthrough clips — three ways in
+
+Three clips are published. Lengths are `ffprobe` readings of the source renders, not estimates:
+
+| Clip | Length | Audio | Journey |
+|---|---|---|---|
+| [NodeRoom — review every agent change](https://youtu.be/3N7sBxFLFOc) | 24s | silent | R1 — the `#story` no-clobber drills |
+| [NodeRoom — from landing to a room](https://youtu.be/qpzHP5-pWvw) | 17s | silent | R2 + R3 — hydrated landing → create a room → join by code |
+| [NodeRoom — the full walkthrough, narrated](https://youtu.be/uvXf7e4hwt4) | 79s | local-TTS voiceover | R1 + R2 + R3 in one continuous pass |
+
+![Fresh-user landing flow: the hydrated landing page loads, Create a room opens the "How should NodeAgent edits land?" policy choice, and the inline ENTER CODE control joins an existing room by share code](docs/walkthroughs/fresh-user-landing.gif)
+<sub>The 17s fresh-user clip inline, because GitHub does not play video in a README. Rendered from the same source MP4 with this repo's two-pass palette recipe ([`scripts/walkthroughs/render.ts`](scripts/walkthroughs/render.ts): `fps=12`, 896px lanczos, `palettegen stats_mode=diff` → `paletteuse` bayer/`diff_mode=rectangle`). Filming it found and fixed a shipped regression: after the Radix migration every `FocusTrapDialog` modal rendered *behind* its own blur scrim. The hydrated React landing also turned out to be a different page from the SSR shell — join is an inline control there, not a dialog.</sub>
+
+<sub>**Coverage, honestly.** These three cover **3 of the 6 NodeRoom journeys** on the journey map (R1 drills, R2 create-a-room, R3 join-by-code); across both products the tally is **10 of 13 journeys shot, 0 reachable and unshot**. The two unfilmed NodeRoom journeys share one root cause, not two: **R5** (in-room review/approve — the product's core journey) and **R4** (mobile approver) both need a signed-in, seeded live room, and R5 is **declined by the owner** ("no seed room") rather than missed. **R6** is not a distinct journey after probing — the `#story` "Architecture" button navigates within the page already filmed in R1. At element level the R1 capture touches **6/21 elements** on an 8,570px surface with 17 controls: a journey clip, not a control sweep.</sub>
+
 ### ▶ Full end-to-end demo — the live analyst room (narrated, with music)
 
 The whole wedge in ~75 seconds — **Capture → Research → Brief → Evidence → Handoff** — with OpenAI TTS
-narration and an original ambient music bed mixed under the voice. This is the only clip here with **audio**.
+narration and an original ambient music bed mixed under the voice. This is the only clip embedded here with **audio**
+(the narrated walkthrough above plays on YouTube).
 
 https://github.com/HomenShum/noderoom/raw/main/episodes/noderoom-analyst-room-v1/renders/short.mp4