docs: record the four end-to-end verification gap checks (run 340)

Add a dated follow-up section to docs/app-integration-verification.md that
closes the four claims the first pass left unproven, each with the command or
capture that proved it against the deployed server (0d1ddc7) on the Pixel 9 Pro
XL 47241FDAS00DLK:

- Chat reply arrives: a throwaway session showed the assistant reply `Done.`
  rendering live, and the close-a-chat flow left the session `closed=true`
  server-side.
- Live pushed run events: disproven on this build. Runs 341/342/343 never
  appeared or advanced without a manual `Refresh runs`; only a local elapsed
  timer ticked. Consistent with the run-event push-hub defect fixed on main by
  cff6f4d, absent from the deployed 0d1ddc7. Recorded as a defect with exact
  reproduction, not an app defect.
- Disconnected state: the closed-port state and the restore to
  ws://10.0.0.75:8080/connect were both captured and Settings verified after.
- Wide windows: the covered-pane region is a constant 553 px at depths 1-3 and
  at 2000x2244 and 2076x2152; `wm size reset` left no override.

Checked for staleness from this change: it touches no source or resource file,
so no packaged template or packaged authoring skill reference
(internal/author/skill/references/) depends on it - none references this
document and none needs updating.

Report-only: the untracked scratch directory .gaps-tmp/ is deliberately left
uncommitted.
This commit is contained in:
2026-10-18 02:46:46 +01:00
parent d0c12e3b75
commit db313336b2
+115
View File
@@ -80,3 +80,118 @@ observed failures.
brief), which the server reports as `canceled`; no build, capture or `adb` process was live and the
Pixel's last capture predated my run. I removed and recreated the lock at 01:23 to take it, and
released it (`rmdir`) when the captures were done. No other lane's work was running on the device.
## Follow-up: the four coverage gaps (2026-10-18, run 340)
This section closes the four items the first pass left open. Same server (`0d1ddc7`, confirmed with
`pipelines ping`), same endpoint `ws://10.0.0.75:8080/connect` with workspace
`/Users/alan/repos/pipelines` (no stub was used and none was left behind), same device
(Pixel 9 Pro XL `47241FDAS00DLK`, 360 dpi, window 1008x2244). The phone app was **not** reinstalled,
so it is the same main build the first pass ended on. Capture directory for every row below:
`/Users/alan/.local/share/pipelines/data/artifacts/340/1/`.
### 1. A chat reply arrives — verified
A throwaway session was created in the app (New chat → Home (~)); never an operator or run session.
Server-side it is `ses_eb3671fa1ffeX7BdCHKfvnE9Kk` (`/Users/alan`, created 2026-10-18T01:20:49Z).
The prompt `reply with one short line` was typed and sent; the composer cleared, the user turn
rendered, and then the assistant reply **`Done.`** rendered in the transcript with no refresh, no
re-navigation and no second send. Evidence: `03b-typed.png` (prompt in the composer),
`04a-reply.png` (user turn + `Done.` in the transcript); server confirmation in
`evidence-chat-messages.json` — `chat_messages` returns the `user` turn `reply with one short line`
and the `assistant` turn `Done.`
Close-a-chat: Chat actions → `Close chat…` → `Close chat` (`06-menu.png`, `07-close-dialog.png`).
The dialog states the chat is marked closed on every device and kept, not deleted. After it the
server reports the session `closed=true`, `closed_at=1792286575309` (2026-10-18T01:22:55Z) —
`evidence-chat-sessions.json` — so the close is real server-side, not a local hide, and reopening the
row renders the same answered turn from history (`60-closed-chat-transcript.png`).
One deviation from the brief's wording, reported as a decision rather than a failure: the closed
chat does **not** disappear from the list. It leaves the active list and moves under a `Closed`
header carrying a `Closed` chip (`08-closed-list.png`); the app deliberately keeps closed chats as
history. The active section no longer contains it and the closed state is server-side, which is the
substance the brief asked for.
### 2. Live pushed run events — disproven on this build (defect)
A real 2-stage run was started from the shell and watched on the phone. The app's run views do
**not** update from pushed frames on the deployed server; every update needed a manual `Refresh runs`.
Reproduction (runs 341/342/343, all created from the shell after `41b-runs-before-new.png`):
- **A new run does not appear in the Runs list.** Run 342 started 01:29:13Z; at 01:29:24Z the list
still lacked it (`41c-list-no-refresh.xml`); one `Refresh runs` tap added it
(`41d-list-after-refresh.xml`). The same held for 341 and 343.
- **A run's view does not advance.** Run 343 (Research a topic) was opened in its run view at
01:30:34Z while stage `research` was running (`42-run-view-t0`,
`51-runview-t0-research-running`). The server recorded it **succeeded, stage `report`, 2 completed
visits, finished 01:30:41Z** (`pipelines inspect 343`). The app kept showing
`running · stage research · 0 completed visits · 1 recorded visit` for the whole ~5-minute watch
(`52-runview-latest`, `53-runview-still`, taken 01:35:56Z) — no stage advance, no terminal
transition, no interaction.
- **The Runs list row was stale the same way.** At 01:36:13Z run 343's row read
`running · 0 completed visits · stage research · run 343 · 6m 13s` (`54-list-before-refresh.png`);
one `Refresh runs` flipped it to `succeeded · 2 completed visits · stage report · run 343 · 18 Oct
02:30 · 20s` (`55-list-after-refresh.png`).
- **The only "live" value is the elapsed clock**, and it is a local timer, not a pushed update:
run 340 counted `14m 41s → 15m 56s → 23m 42s` with no interaction, while run 343's clock reached
`6m 13s` although the server had finished it at `20s`.
Cross-checks that rule out a dead socket: chat streamed a reply live at 01:21 in the same app session
(item 1), and Settings read `Connected` (`32-settings-endpoint-restored`). The server log shows the
app's single `subscribe` (android-5, 01:26:47Z), then only `list_runs` on each manual refresh and
`get_run` on each open — no further subscription traffic. This is consistent with the known
run-event push-hub defect the in-flight run 339 fixes ("an oversized subscribe snapshot overflow"),
which is not in the deployed `0d1ddc7` build; it is not an app defect, and it is exactly what the
brief's item 2 tested for. **Not verified, and disproven as stated:** run state does not arrive by
push on this deployment; a manual refresh is required.
### 3. Disconnected state — verified
One command sequence deviated and restored (`.gaps-tmp/disconnected.py`, run 340's scratch): set the
URL to the closed local port `ws://127.0.0.1:9/connect`, tapped Connect, captured, then set the real
URL back, tapped Connect, force-stopped and relaunched. The disconnected state read
`Disconnected: Failed to connect to /127.0.0.1:9 · retrying with backoff`
(`30-disconnected.png`/`.xml`). After the restore the app showed live server data again — run 340
`running 14m 10s` (`31-restored.png`/`.xml`) — and Settings confirmed
`ws://10.0.0.75:8080/connect`, `Connected`, build `0d1ddc7`, workspace `/Users/alan/repos/pipelines`
(`32-settings-endpoint-restored`). The phone was never left on a dead endpoint.
### 4. Wide-window checks — verified
`adb … shell wm size 2000x2244` and `wm size 2076x2152` (the fold's shape), then `wm size reset`. At
360 dpi the covered region (the peek strip, i.e. the active pane's left edge) was read from the
covered pane's own tap-target bounds in `uiautomator dump`:
| Depth | Covered edge bounds (px) | Covered region |
| --- | --- | --- |
| 1 (pipeline detail) | `[67,176,553,…]` | 553 px |
| 2 (stage pane) | `[67,176,175,…]` + `[175,176,553,…]` | 553 px |
| 3 (stage section) | `[67,176,175,…]` + `[175,176,553,…]` | 553 px |
At 2076x2152 the region is the same 553 px at depths 1 and 3. So **depth adds no width**: the region
is identical at depth 1, 2 and 3 and at both window sizes. 553 px = 245.8 dp = the measured
compact-identity width (216 dp = `MIN_COVERED_PANE_WIDTH_DP` 112 + `PEEK_WIDTH_DP` 56 + one 48 dp
touch target — the full-statistics budget) plus the 30 dp leading gesture inset. The older edge is
175 px = one touch target after the inset, so no covered edge is squeezed below the tappable
minimum; the measured width sits on the intended budget, inside the allowed band. Evidence:
`21-wide-pipelines-list.png`, `22-wide-depth1.png`, `24-wide-depth2.png`, `25-wide-depth3.png`,
`26-fold-depth3.png`, `27-fold-depth1.png`. After the reset `wm size` reports only
`Physical size: 1008x2244` and `dumpsys window displays` reports `cur=1008x2244` — no override left
(`28-after-reset`).
### Second client / fold
The brief allowed the fold's own view only if it cost no more than a capture; it was not opened. The
shared-state proof is instead a different client: the CLI, querying the same server, sees the created
session `closed=true` (`evidence-chat-sessions.json`) — the registry state the protocol says every
device shares.
### Still unverified
The fold's own client view (not opened) and the chat streaming's intermediate frames (the reply was
captured complete, not frame by frame). Everything else across the four gaps is settled above.
Device lock: taken at 02:13 local (`mkdir …/device.lock`) and released with `rmdir` when the captures
were done; no other lane's device work was running.