# Evidence for [#26794](https://git.cafe/pingdotgg/t3code/pull/26794)

> **Revision v2 (lead round 2):** R01–R06 transcripts re-uploaded as `*-sanitized.jsonl` with the host's provider skill descriptions replaced by a placeholder ([sanitize-protocol.py](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/c623540d9301299f/sanitize-protocol.py)); line numbers and all protocol data are unchanged. R10 recaptured with schema-valid fixtures. The v1 R10 transcripts and the unsanitized protocol files are superseded.

Beyond faster warm connections, the head changes status process coalescing and publishes late network snapshots. Those additional observable changes are included below. The boot-overlap wait remains; this evidence does **not** claim a 250 ms end-to-end limit.

Captured **11 / 11 affected rows**: R01–R10 and W01. U01/U02 are unchanged branch inventories with no mandatory extra capture. Unverified required rows: **none**.

Before: [fe219c4](https://github.com/pingdotgg/t3code/commit/fe219c4de18f5ee37c28cbd3914a5569b1579156); after: [4ecd25f](https://github.com/pingdotgg/t3code/commit/4ecd25f21cf8f3755fff58967daad1fae22a74fd). Both ran as the real source application from separate pinned checkouts. No repository files were edited, no tests replaced captures, and no PR was edited, commented on, pushed or committed.

Rendering environment (W01): Playwright-core 1.60.0, Chromium 148.0.7778.96, Linux x64 headless, no device emulation, DPR 1, font scaling 100%, viewport/video 1440×900, dark. The primary and web bundle are the same head checkout for both target sides. Server rows: viewport/theme N/A.

Real-server controls: fresh per-run database created at that side's migration level; `T3CODE_DEV_SEED=0`; explicit data directory under scratch; loopback bind; real sshd already listening on IPv4 and IPv6 port 22; unchanged host/interfaces across sides. Published host/user labels are synthetic (`fake-host`, `synthetic-user`). B0=[] for the deliberate loopback bind; M=`fake-host.local`. No live database was read or copied. The standalone CLI calls the explicit data flag `--base-dir`; the dev runner calls it `--home-dir`.

The fake CLI latches mode at spawn, logs every start/finish/SIGTERM with PID and Unix UTC timestamp, handles Serve and prefs immediately, and gates selected status outcomes. Slow-failure status has a 10-second planned delay and is terminated by production's unchanged 3.5-second timeout. HTTPS self-probes to fake DNS fail in real runs; no real HTTPS endpoint is claimed. Positive HTTPS evidence in R07/R08/R09 is explicitly a controlled HttpClient-200 service harness.

All real transcripts contain ticket request/response, WS open, Request/Chunk/ACK/Exit, valid first snapshots, follow-ups, interruption/close and process cleanup. Raw CLI logs retain startup and setup attempts. Use the UTC row boundaries, not whole-file totals, for row spawn counts. Some `snapshot.valid.ticketToSnapshotMs` values are cumulative from that subscriber's ticket even on follow-ups.

## R01 — actual connect during boot read

Steps: start fresh Serve-on server; connect immediately on its Listening log; subscribe and ACK; wait for process settlement and issue a `server.probe` fence. Head boot status PID starts 734.91 ms **before** the ticket request and remains pending until its timeout. Baseline has no eager boot status process. This row used real startup successfully; no boot harness fallback was needed.

Before scripted row: 2026-10-29T06:13:32.786Z → 2026-10-29T06:13:39.741Z. After scripted row: 2026-10-29T06:13:34.137Z → 2026-10-29T06:13:36.922Z.

| Before | After |
|---|---|
| [fe219c4](https://github.com/pingdotgg/t3code/commit/fe219c4de18f5ee37c28cbd3914a5569b1579156) | [4ecd25f](https://github.com/pingdotgg/t3code/commit/4ecd25f21cf8f3755fff58967daad1fae22a74fd) |
| Ticket→first snapshot **6954.05 ms**; 2 uncached status processes; M/B0. No boot status process existed when the request began. | Ticket→first snapshot **2784.62 ms**; 1 boot status process; M/B0. `previous` is unknown because the only boot PID has not settled at request time. No distinct target snapshot follows identical re-resolution. |
| [before-real-protocol-sanitized.jsonl](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/bd5338e954cf14fc/before-real-protocol-sanitized.jsonl) | [after-real-protocol-sanitized.jsonl](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/849c973859c446ae/after-real-protocol-sanitized.jsonl) |

## R02 — warm Serve-on burst and reconnect

Steps: after R01 settles, open A and B together; observe status spawn; cancel A while pending; open replacement; await survivor/replacement snapshots; make unary `server.getConfig` while subscription remains open; await status terminal markers; use `server.probe` as fence. The cancelled A intentionally has no snapshot.

Before scripted row: 2026-10-29T06:13:39.747Z → 2026-10-29T06:13:53.023Z. After scripted row: 2026-10-29T06:13:36.951Z → 2026-10-29T06:13:40.507Z.

| Before | After |
|---|---|
| [fe219c4](https://github.com/pingdotgg/t3code/commit/fe219c4de18f5ee37c28cbd3914a5569b1579156) | [4ecd25f](https://github.com/pingdotgg/t3code/commit/4ecd25f21cf8f3755fff58967daad1fae22a74fd) |
| B **6589.75 ms**, replacement **6572.75 ms**; **7** status spawns in row; one snapshot per surviving subscription; M/B0. | B **328.58 ms**, replacement **280.57 ms**; **1** shared status spawn in row, surviving cancellation and reconnect; one snapshot per survivor; M/B0; no failure→failure network follow-up. |
| [before-real-protocol-sanitized.jsonl](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/bd5338e954cf14fc/before-real-protocol-sanitized.jsonl) | [after-real-protocol-sanitized.jsonl](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/849c973859c446ae/after-real-protocol-sanitized.jsonl) |

## R03 — warm Serve off

Steps: separate Serve-off run; warm-up subscription; connect and reconnect sequentially while the slow status plan is unchanged; await settlement; fence. The captured off-run driver used `server.getConfig` for its final fence, so the count includes that additional baseline refresh. Its exact behavior is retained in capture-off-recorded.mjs. No Serve CLI invocation occurs.

Before scripted row: 2026-10-29T06:12:00.988Z → 2026-10-29T06:12:15.533Z. After scripted row: 2026-10-29T06:12:02.131Z → 2026-10-29T06:12:08.478Z.

| Before | After |
|---|---|
| [fe219c4](https://github.com/pingdotgg/t3code/commit/fe219c4de18f5ee37c28cbd3914a5569b1579156) | [4ecd25f](https://github.com/pingdotgg/t3code/commit/4ecd25f21cf8f3755fff58967daad1fae22a74fd) |
| Connect **3553.71 ms**, reconnect **3563.64 ms**; **3** status spawns for measured connect/reconnect/unary-fence (2 for the two subscriptions alone); M/B0. | Connect **286.54 ms**, reconnect **280.93 ms**; **1** shared status spawn across measured calls; M/B0. |
| [before-real-protocol-sanitized.jsonl](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/bd5338e954cf14fc/before-real-protocol-sanitized.jsonl) | [after-real-protocol-sanitized.jsonl](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/849c973859c446ae/after-real-protocol-sanitized.jsonl) |

## R04 — fresh success fits grace

Steps: one staged fast run covers successful boot/read, F→A, and A→B; retain A subscription through transitions. Then the transitions run explicitly changes A→B using unary `server.getConfig` while an older subscriber is open and opens another B subscriber. Fast process start/finish records establish that status itself settled inside grace; total config time also contains transport/editor/HTTPS-probe work.

Before scripted row: 2026-10-29T06:12:47.122Z → 2026-10-29T06:12:47.865Z. After scripted row: 2026-10-29T06:12:45.840Z → 2026-10-29T06:12:46.887Z.

| Before | After |
|---|---|
| [fe219c4](https://github.com/pingdotgg/t3code/commit/fe219c4de18f5ee37c28cbd3914a5569b1579156) | [4ecd25f](https://github.com/pingdotgg/t3code/commit/4ecd25f21cf8f3755fff58967daad1fae22a74fd) |
| Fast-run initial A **398.83 ms**, F→A **116.83 ms**, A→B **108.98 ms**; **8** status spawns across fast run. Unary companion: 6 status spawns; old subscriber stays at 1 snapshot, new B **156.73 ms**. | Fast-run initial A **539.85 ms**, F→A **147.80 ms**, A→B **218.88 ms**; **5** status spawns including eager boot. Unary companion: 3 status spawns; old subscriber receives B, 2 snapshots total; new B **80.74 ms**. Initial targets are fresh T(A)/T(B), no duplicate network-only initial payload. Real H remains absent. |
| [before-real-protocol-sanitized.jsonl](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/bd5338e954cf14fc/before-real-protocol-sanitized.jsonl) | [after-real-protocol-sanitized.jsonl](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/849c973859c446ae/after-real-protocol-sanitized.jsonl) |

## R05 — fast failure and failure equivalence

Steps: keep the R04 successful B observer open (B is the seeded successful name for this row); switch to malformed JSON/exit 0 and open a subscriber; switch to nonzero exit and open another; await terminal records and probe fences after each. A true missing-executable variant was optional and was not run.

Before scripted row: 2026-10-29T06:19:19.977Z → 2026-10-29T06:19:20.274Z. After scripted row: 2026-10-29T06:19:20.842Z → 2026-10-29T06:19:20.992Z.

| Before | After |
|---|---|
| [fe219c4](https://github.com/pingdotgg/t3code/commit/fe219c4de18f5ee37c28cbd3914a5569b1579156) | [4ecd25f](https://github.com/pingdotgg/t3code/commit/4ecd25f21cf8f3755fff58967daad1fae22a74fd) |
| Malformed **129.73 ms**, nonzero **159.20 ms**; **4** status spawns; both new snapshots M/B0. Old subscriber remains at 1 initial snapshot because baseline has no network follow-up path. | Malformed **82.10 ms**, nonzero **60.37 ms**; **2** status spawns. Both new snapshots immediately drop T(B). Old observer count becomes 3 (A, B, then M-only) after success→failure and stays 3 after the different failure. |
| [before-real-protocol-sanitized.jsonl](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/bd5338e954cf14fc/before-real-protocol-sanitized.jsonl) | [after-real-protocol-sanitized.jsonl](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/849c973859c446ae/after-real-protocol-sanitized.jsonl) |

## R06 — late recovery, ACK buffer, cancellation

Steps: settled F; latch a gated A status plan; open two subscribers; cancel one after the status-start marker; withhold survivor ACK; hold gate for 1002 ms before / 1001 ms after; configure the hypothetical next status to fail; release gate; await status terminal records; release survivor ACK; await recovered snapshot and probe fence. Gates and mode changes are explicit in the transcript.

Before scripted row: 2026-10-29T06:19:20.274Z → 2026-10-29T06:19:21.410Z. After scripted row: 2026-10-29T06:19:20.993Z → 2026-10-29T06:19:22.185Z.

| Before | After |
|---|---|
| [fe219c4](https://github.com/pingdotgg/t3code/commit/fe219c4de18f5ee37c28cbd3914a5569b1579156) | [4ecd25f](https://github.com/pingdotgg/t3code/commit/4ecd25f21cf8f3755fff58967daad1fae22a74fd) |
| First/recovered snapshot **1130.71 ms**; 1 snapshot containing T(A)+M/B0; **3** status spawns including cancelled caller and uncached discovery calls. | Stale first snapshot **298.30 ms**, recovered follow-up **1188.50 ms** cumulative from ticket; 2 snapshots; **1** status spawn. Refresh survives cancellation, queued recovery crosses the ACK boundary, and settled-only re-resolution does not start the configured failing next call. |
| [before-real-protocol-sanitized.jsonl](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/bd5338e954cf14fc/before-real-protocol-sanitized.jsonl) | [after-real-protocol-sanitized.jsonl](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/849c973859c446ae/after-real-protocol-sanitized.jsonl) |

## R07 — late loss of remote and HTTPS targets

**LABELLED deterministic service + production accumulator harness.** Steps: real reader/parser/resolvers seed A with controlled SSH and HttpClient 200; latch gated failure; advance TestClock 300 ms; release at 1000 ms; await reader publication and settled-only re-resolution; run the production accumulator to completion; replay identical failure. This is not a claim that the fake DNS was reachable on the VM.

| Before | After |
|---|---|
| [fe219c4](https://github.com/pingdotgg/t3code/commit/fe219c4de18f5ee37c28cbd3914a5569b1579156) | [4ecd25f](https://github.com/pingdotgg/t3code/commit/4ecd25f21cf8f3755fff58967daad1fae22a74fd) |
| Gated paired resolver workload waits until 1000 ms and initially returns M/B0; **2** status handles for transition; no baseline late network path or fabricated follow-up. | At 300 ms first payload is T(A)+M and H(A); at ~1000 ms publication/re-resolution produces 1 full snapshot removing both T(A) and H(A), retaining M/B0; **1** status handle for transition, no recursive refresh. Repeated failure publishes no additional reader change. |
| [before-service-harness.jsonl](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/0c053127e5fa6020/before-service-harness.jsonl) | [after-service-harness.jsonl](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/04e373125700797f/after-service-harness.jsonl) |

## R08 — late successful rename

**LABELLED deterministic service + production accumulator harness**, shared session with R07. Steps: seed A including verified-in-harness H(A); latch gated B; inspect at 300 ms; release at 1000 ms; drain reader change/resolution/accumulator; repeat identical B. Exact virtual clock values for the second stage appear in the transcript.

| Before | After |
|---|---|
| [fe219c4](https://github.com/pingdotgg/t3code/commit/fe219c4de18f5ee37c28cbd3914a5569b1579156) | [4ecd25f](https://github.com/pingdotgg/t3code/commit/4ecd25f21cf8f3755fff58967daad1fae22a74fd) |
| Initial paired resolver payload waits for B and includes T(B)+M/H(B); **2** transition status handles; 0 late network snapshots (mechanism absent). | First payload at 300 ms retains T(A)+M/H(A); one late snapshot replaces both with T(B)+M/H(B), with no accumulated duplicate host; **1** transition status handle. Repeated B adds no reader publication; total transition+repeat handles 2. |
| [before-service-harness.jsonl](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/0c053127e5fa6020/before-service-harness.jsonl) | [after-service-harness.jsonl](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/04e373125700797f/after-service-harness.jsonl) |

## R09 — reader changes, target projection unchanged

**LABELLED deterministic reader/resolver + production accumulator harness.** Two variants: keep DNS A and probes fixed while changing only parsed tailnet IPv4 `.10`→`.99`; and successful no-DNS→failure. Each gates the change for 1000 virtual ms, awaits the head reader publication, re-resolves with SettledReadOnly, and runs the finite accumulator stream to completion. Identical outcome replay and TestClock drainage establish the negative emission result without arbitrary silence.

| Before | After |
|---|---|
| [fe219c4](https://github.com/pingdotgg/t3code/commit/fe219c4de18f5ee37c28cbd3914a5569b1579156) | [4ecd25f](https://github.com/pingdotgg/t3code/commit/4ecd25f21cf8f3755fff58967daad1fae22a74fd) |
| Each paired resolver transition uses **2** status handles, 0 network follow-up snapshots; baseline has no shared reader publication channel. | Each transition uses **1** status handle and produces **1 reader publication / completed late resolution**, but **0 network snapshots** because both target arrays are unchanged. Identical replay adds 0 reader publications; queue is empty at completion. Reader equality and target equality are separately recorded. |
| [before-service-harness.jsonl](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/0c053127e5fa6020/before-service-harness.jsonl) | [after-service-harness.jsonl](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/04e373125700797f/after-service-harness.jsonl) |

## R10 — late full snapshot preserves newer config (v2)

**LABELLED deterministic harness:** production `withLateConfig` accumulator and client projection, plus production `RpcServer.make` with JSON serialization over an in-memory protocol. Recaptured as v2 because the v1 capability companion sent a theme without `appearance` and a usage source without `kind`/`checkedAt`/`accounts`, which production decoding rejected (an unsolicited Failure/Die). Details, steps and scripts: [R10-v2.md](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/0689852630c71d43/R10-v2.md), [fixtures-v2.mjs](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/492f83e0c7bb9024/fixtures-v2.mjs).

| Before | After |
|---|---|
| [fe219c4](https://github.com/pingdotgg/t3code/commit/fe219c4de18f5ee37c28cbd3914a5569b1579156) | [4ecd25f](https://github.com/pingdotgg/t3code/commit/4ecd25f21cf8f3755fff58967daad1fae22a74fd) |
| 4/4 accumulator scenarios: 1 late editor snapshot each; newer settings, keybindings, issues, providers, themes and sources survive; no late network path exists (recorded as absent). RPC: 0 Failure/Die; every enabled theme/source event decoded and projected. 0 status processes. | 4/4 accumulator scenarios: 2 late snapshots each (editor and network), newer state preserved in both orderings. RPC: 0 Failure/Die; every enabled theme/source event decoded and projected. 0 status processes. |
| [before-R10-harness-v2.jsonl](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/07f4508d6f9f4e3d/before-R10-harness-v2.jsonl) | [after-R10-harness-v2.jsonl](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b3966145e8ac9858/after-R10-harness-v2.jsonl) |

## W01 — real web Connecting → Connected recording

Steps and timestamps above the pair: saved **Evidence target** starts switched off; click its enable switch; keep the same view until Connected. Target boot has already settled. Before: click `06:18:50.385Z`, Connecting `06:18:50.583Z`, Connected `06:18:57.485Z`. After: click `06:18:51.675Z`, Connecting `06:18:51.807Z`, Connected `06:18:52.396Z`. Timings are 2026-10-29 UTC; monotonic browser timestamps are retained in the transcript. Videos omit the initial page-loading/setup phase, retain the off state before the click, and have a top revision label.

| Before | After |
|---|---|
| [fe219c4](https://github.com/pingdotgg/t3code/commit/fe219c4de18f5ee37c28cbd3914a5569b1579156) | [4ecd25f](https://github.com/pingdotgg/t3code/commit/4ecd25f21cf8f3755fff58967daad1fae22a74fd) |
| Click→Connected **7104 ms**; visible Connecting **6896.70 ms**; **2** status spawns during measured connect | Click→Connected **725 ms**; visible Connecting **581.50 ms**; **1** status spawn during measured connect |
| [W01-before.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/1a7181b65889319b/W01-before.mp4) | [W01-after.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/082077d18f06a785/W01-after.mp4) |

The raw protocol and browser event timelines are in the real-protocol files. The same head web bundle/primary and fresh saved-route state are used on both sides. Different target loopback ports (43881 vs 43882) identify the sides. No bearer token or pairing URL appears in the recording.

No settled Connected stills are included, so still diff images and still pixel-stability captures do not apply. Review contact sheets stayed in scratch; both final 1440×900 videos were opened and visually inspected, and then played through `ended` in Chromium with no decode error: [recording-review.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/6c2e00d5241d823f/recording-review.json).

## What else changed

- Head adds eager boot status discovery (including Serve-off) and shares a pending read. R01 still waits for the boot read's remaining timeout; it is not grace-bounded.
- Warm first-snapshot latency falls, while the WebSocket open itself is fast on both sides. The 250 ms value bounds the warm reader wait, not total ticket→snapshot or visible readiness.
- Status spawn counts fall in the matched real row windows: R02 7→1; R03 measured calls 3→1; fast R04 whole run 8→5 (head count includes boot); unary R04 companion 6→3; R05 4→2; R06 3→1; W01 2→1. Fast staggered reads can still launch multiple processes.
- Existing subscribers now receive full network snapshots after successful recovery, loss, rename, and fast changes driven by another subscriber or unary config request. Baseline has no such network follow-up path. It continues receiving its existing editor updates.
- Failure→failure, identical success, and status-only changes with equal target arrays do not churn network snapshots. R09 distinguishes the reader publication from the suppressed network output.
- Late network/editor snapshots preserve newer settings, keybindings/issues, providers and reveal kind. Themes and usage sources remain independent events and survive client projection.
- No changed target projection appears in the warm failure cases: both retain M and B0. Real fake-DNS probes never add H; only the labelled harness advertises positive H after a controlled 200.
- Nonfunctional fixture differences are target ports, isolated home/checkout paths, generated environment/session/cookie IDs, provider checkedAt timestamps and elapsed time in the blurred background. Settings values in the matched real R02 snapshots are equal. The target route label, viewport, theme and layout are the same. Enabling a route expands the existing load-balancing/sharing rows on both sides; this is existing client behavior.
- Whole-file raw CLI totals include setup/retakes and later staged rows (core file totals 26 before / 10 after), so they must not be mistaken for row-window counts. All original terminal records remain present.

## Rebuild the state

1. Recreate the detached baseline at the pinned SHA with the worktree command in the brief and run global `vp install`. Keep the provided after checkout at the pinned head. No source edits or builds outside the source runtime are needed.
2. Put the supplied scripts under `/tmp/lead-26794-r2/evidence-capture/` and make `tailscale` executable. It is prepended to PATH only in spawned test processes; global vp/Node selection is unchanged.
3. First run the standalone Serve-off pair with `capture-off-recorded.mjs <side> off` and fast pair with `capture.mjs <side> fast` while the corresponding target ports are free. These staged runs each use one server start per side.
4. Run `capture.mjs before onweb` and `capture.mjs after onweb`. The script issues an isolated bearer through the real CLI, starts real `t3 serve`, captures R01/R02, and leaves the targets available for W01/transitions. Its `*.private.json` handoff contains a fresh token and must stay private.
5. Start `web-primary.mjs`; it invokes `T3CODE_DEV_SEED=0 vp run dev --home-dir <scratch>/home --no-browser`, records actual dev-runner ports privately, and uses a fast failing fake CLI. Run `browser-inspect.mjs` once to pair that private primary and save its cookie state. Then run `record-w01.mjs before` and `record-w01.mjs after`. It seeds the route through the app's catalogue commands using schema classes, switches it off, waits for prior status termination, and records the enable interaction. Run `transitions.mjs <side>` against the held targets for the unary R04 companion and R05/R06. These stages share the original servers.
6. `service-harness.mjs <side> R09` and `<side> R07-R08` run the real parser/reader/resolvers with simulated child handles, controlled sshd and HTTP 200, Deferred gates and TestClock. `accumulator-harness.mjs <side>` drives real production accumulator and client projection. `build-rpc-harness.py` builds the standalone RPC companion from existing test scaffolding without running a test; run each generated `rpc-harness.mts` to capture actual default and enabled subscription payloads. The generated source is uploaded too.
7. Signal `web.complete` in each held target's run directory so its owning driver stops its tracked server. Stop the primary's captured process group. Remove the baseline worktree. The capturer did all of this; [cleanup-audit.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/9106054586ee130e/cleanup-audit.json) confirms all 10 tracked server launches are stopped, browsers closed and the baseline removed. Scratch is retained for the lead.

All drivers create fresh timestamped run directories. No raw logs or previous takes are deleted. The final source-runtime CLI is the real app; a separate bundle build was unnecessary. No paid providers or turns were used.

## Artifact inventory

- [W01-after.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/082077d18f06a785/W01-after.mp4)
- [W01-before.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/1a7181b65889319b/W01-before.mp4)
- [accumulator-harness.mjs](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/73fa1e484ac07af0/accumulator-harness.mjs)
- [after-R10-harness.jsonl](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/c1a1f3aa3c716e23/after-R10-harness.jsonl) (v1, superseded)
- [after-cli.jsonl](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/5d9d4c669d48e3fb/after-cli.jsonl)
- [after-real-protocol-sanitized.jsonl](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/849c973859c446ae/after-real-protocol-sanitized.jsonl)
- [after-server.log](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/47c58ffd3eef7615/after-server.log)
- [after-service-harness.jsonl](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/04e373125700797f/after-service-harness.jsonl)
- [before-R10-harness.jsonl](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/0e064284f0ef84ac/before-R10-harness.jsonl) (v1, superseded)
- [before-cli.jsonl](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/1072dc781d98560b/before-cli.jsonl)
- [before-real-protocol-sanitized.jsonl](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/bd5338e954cf14fc/before-real-protocol-sanitized.jsonl)
- [before-server.log](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/95aceb322c6ae5fa/before-server.log)
- [before-service-harness.jsonl](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/0c053127e5fa6020/before-service-harness.jsonl)
- [browser-inspect.mjs](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b30f8e142f8d486a/browser-inspect.mjs)
- [build-rpc-harness.py](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/440b42f6451676a5/build-rpc-harness.py)
- [build-transitions.py](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/a8020c07c1c5dd37/build-transitions.py)
- [capture-off-recorded.mjs](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/a475018158370c23/capture-off-recorded.mjs)
- [capture.mjs](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/57e1fc0f216dc84f/capture.mjs)
- [cleanup-audit.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/9106054586ee130e/cleanup-audit.json)
- [generate-index.py](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/8646a5eceb90a269/generate-index.py)
- [matrix.md](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/06a52e93f343e7b3/matrix.md)
- [package-evidence.py](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/6759082706faf476/package-evidence.py)
- [record-w01.mjs](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/834ad7c4e35da1eb/record-w01.mjs)
- [recording-review.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/6c2e00d5241d823f/recording-review.json)
- [review-recordings.mjs](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/a4bc06173d8d1ea7/review-recordings.mjs)
- [review-videos.py](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/da5a57e01baa0b22/review-videos.py)
- [rpc-after-harness.mts](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/3c9d9a2c8d42287f/rpc-after-harness.mts)
- [rpc-before-harness.mts](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/0b3318507bc248f9/rpc-before-harness.mts)
- [service-harness.mjs](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/9c83efd8cf7b4353/service-harness.mjs)
- [tailscale](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/e0c223312d404a32/tailscale)
- [transitions.mjs](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/a3ab36ad4070375d/transitions.mjs)
- [trim-videos.py](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/ed3e68f2c8c8bfed/trim-videos.py)
- [upload-artifacts.py](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/160dfacb28eec316/upload-artifacts.py)
- [web-primary-server.log](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/705f069cabca734f/web-primary-server.log)
- [web-primary.mjs](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/667e44e8d7ddd9a7/web-primary.mjs)

## Friction

- tool | agent-browser was absent; the brief-authorized Playwright path worked | preflight the browser CLI.
- brief | standalone `t3 serve` exposes `--base-dir`; `--home-dir` is the dev-runner spelling | distinguish the two command flags.
- model | initial W01 fixture used plain objects instead of schema classes and an exact-text status selector; setup was corrected and the recording retaken | seed with the app's tagged classes and wait on row state.
- tool | import-only workspace exports cannot be resolved with CJS require.resolve; harnesses use source imports or the checkout's normal ESM module resolution | use ESM resolution for workspace packages.

Completed within the 70-minute capture cutoff. No required row was omitted.

### Added in v2

- [sanitize-protocol.py](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/c623540d9301299f/sanitize-protocol.py)
- [R10-v2.md](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/0689852630c71d43/R10-v2.md)
- [before-R10-harness-v2.jsonl](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/07f4508d6f9f4e3d/before-R10-harness-v2.jsonl)
- [after-R10-harness-v2.jsonl](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b3966145e8ac9858/after-R10-harness-v2.jsonl)
- [fixtures-v2.mjs](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/492f83e0c7bb9024/fixtures-v2.mjs)
- [accumulator-harness-v2.mjs](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/035b1300532cab7c/accumulator-harness-v2.mjs)
- [build-rpc-harness-v2.py](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b34f64a5d788e1d4/build-rpc-harness-v2.py)
- [rpc-before-harness-v2.mts](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/e15d19b3f59b9e05/rpc-before-harness-v2.mts)
- [rpc-after-harness-v2.mts](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/dbe6e09e45dd6196/rpc-after-harness-v2.mts)
- [run-r10-v2.py](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/7d1c736df24a4ecc/run-r10-v2.py)