659 S4 merge train — CI remediation
# #659 S4 merge train — CI remediation
Initial source: `a1b089cac0b8cf752b2aec4453eae8b8c6ce4c0b` (PR #13).
PC-A worker `task_5f56b60d7cf1` / `ctx_efa1cf97bbff`, 2026-10-11 KST.
The verified stack remains linear, main `9b7e1c31` through PR #2–#13,
15 commits and 899 changed files, with zero baseline PNG changes.
## First same-head runs
[CI attempt 2](https://github.com/coco-de/demo-ecommerce-grocery/actions/runs/38042991771/attempts/2)
completed with 11 successful jobs and three failures.
| Job | Result |
|---|---|
| 변경 범위 | success |
| 계약 게이트 | success |
| 번들 모듈 → 호스트 코드 생성 | success |
| backend 통합 (생성 SDK → 호스트 → 격리 Postgres) | success |
| Android 빌드 (debug) | success |
| 검증 (tests-modules) | success |
| 검증 (analyze) | success |
| 검증 (generated) | failure |
| 검증 (tests-2) | success |
| 검증 (tests-server) | success |
| 검증 (bdd) | success |
| 검증 (widgetbook) | failure |
| 검증 (tests-1) | failure |
| 검증 (web) | success |
The generated job regenerated cleanly, then its format check identified four
co_bdd Patrol outputs in the app/console integration-test directories. The
consuming repository's official `flutter-codegen` command now applies Dart
formatting to these outputs after generation. The committed output is produced
by tools; generated source, SDK imports and migration history are never edited
by hand. CI format exclusions and supplier references are unchanged.
Widgetbook and tests-1 terminated with a runner shutdown signal and a cancelled
step before their configured limits. Widgetbook's one permitted job retry,
[job 114277279425](https://github.com/coco-de/demo-ecommerce-grocery/actions/runs/38042991771/job/114277279425),
repeated the same termination. Times are measured from the GitHub job/step API:
| Execution | Test seconds / limit | Job seconds / limit |
|---|---|---|
| Widgetbook attempt 2 | 379 / 600 | 479 / 900 |
| tests-1 attempt 2 | 338 / 600 | 430 / 900 |
| Widgetbook attempt 3 | 397 / 600 | 493 / 900 |
These results are failures. They are not timeout evidence. The progressive
slowdown and the local memory measurements below support memory pressure.
The coordinator authorized preserving every case while bounding file/isolate
and image-cache lifetimes; no repeated unchanged infrastructure retry is used.
[Pages attempt 2](https://github.com/coco-de/demo-ecommerce-grocery/actions/runs/38042991710/attempts/2)
passed its build, credential scan and pre-deployment smoke: eight checks passed,
six surfaces returned and reloaded with HTTP 200, and console/page errors,
external/business requests and broken requests were all zero. Deployment and
the public smoke jobs were skipped because this was a PR run.
[Golden attempt 2](https://github.com/coco-de/demo-ecommerce-grocery/actions/runs/38042991636/attempts/2)
was cancelled while compare was still running after the repeated CI failure.
The always-run gallery capture required a force-cancel request to stop it.
No full compare, golden difference CSV, baseline update or bot commit is claimed.
## Bounded Widgetbook rendering
The previous file registered all 21 business pages × 11 locales × two themes
in one isolate. Each test still renders data/loading/empty/error, verifies
locale, theme, RTL/LTR and absence of exceptions, and exercises error → retry
→ data. Its callback tokens, including strings and assertions, are identical
after extraction into `test/support/widgetbook_page_matrix.dart`.
The same 462 matrix cases now run through 22 locale/surface entry files. Each
loads the same shipped fonts and the locale it needs. Decoded/live images are
cleared between fixtures. The first package-owned `dart_test.yaml` trial ran one file at a time
so native font caches disappear with each test process. The final measured
concurrency setting is recorded below. This configuration
applies to both the Widgetbook job and the ordinary test shard. It adds no
skip, sampling, retry, error suppression or selection opt-out.
Same command for both local measurements:
```sh
PATH=/Users/dongwoo/fvm/versions/3.47.6/bin:$PATH \
TEST_SUITE=widgetbook /usr/bin/time -l bash tool/dev/test_all.sh
```
A five-second process-tree sampler tracked only descendants of this command.
| Measurement | Before | After |
|---|---|---|
| Result | aborted at disk guard; not PASS | exit 0 |
| Elapsed | 349.65 s, incomplete | 355.18 s |
| `time -l` maximum RSS | unavailable after SIGTERM | 2,363,129,856 bytes (2.20 GiB) |
| Sampled process maximum | 7.39 GiB | 2.18 GiB |
| Sampled process-tree maximum | 7.45 GiB | 3.32 GiB |
| Test result | incomplete | 525 passed, 0 failed, 0 skipped |
| Matrix inventory | source plan 462 | 462 unique; 42 per locale |
The baseline guard stopped this worker's process group when available disk fell
below 40 GiB. The coordinator recovered disk space before the after measurement;
the after run stayed above 45.8 GiB. The baseline's missing `time -l` footer is
not substituted with a successful run. Sampled RSS is recorded separately from
the time tool's maximum. Local macOS measurement is not Linux golden evidence.
## Validation and next checkpoint
- Widgetbook: 525/525 passed, no skip; matrix callback and full inventory preserved.
- Gate regression tests: 88 passed, including bundle/layer/Docker checks.
- Analysis: no issues under `dart analyze --fatal-infos`.
- Official Flutter regeneration: exit 0 in 21.989 s; final bytes of all 1,358
fingerprinted Dart/pubspec/lock files unchanged, including normalized Patrol.
Durable raw measurements, case inventory, JSON events, API job metadata and
sanitized logs are under PC-A
`/Users/dongwoo/orca/toolchains/637/xhost/s4-659-merge-train/`.
The next checkpoint is the completed formatter/memory commit on PR #13 followed
by full CI at that head, exhaustive Linux golden classification, one Linux
baseline update and all green checks at the final bot head before merge.
Supplier baseline v1 remains co-package `c003ca2c`, coui v0.158.0 and Serverpod
4.0.3. Existing live locale/full route selection blockers, Mobbin access,
verified bundle profiles and the approved Patrol runner remain open. W4-12 and
W4-13 remain outside this worker's assignment.
## Hosted serial-run time budget
At remediation head `45a4e32`, [CI run 38077035275](https://github.com/coco-de/demo-ecommerce-grocery/actions/runs/38077035275) passed twelve jobs, including generation/format, and failed only tests-1 and Widgetbook. Both log an explicit ten-minute step timeout; each step lasted 613 seconds including shutdown cleanup. The case matrix kept progressing through separate files, so this is distinct from the original early runner shutdown.
The coordinator-approved change gives only these two steps twenty minutes and only their matrix jobs thirty minutes for preparation and cleanup. Other jobs and steps, the shard assignment, test cases and assertions retain their existing settings. `actionlint` validates all three workflows.
## Measured file concurrency
The coordinator authorized concurrency two only if its sampled process-tree peak stayed at or below 5 GiB. The second full measurement passed the same 525 case names, with zero failures/skips, in 233.86 seconds. Peak tree RSS was 4.971 GiB, and `time -l` maximum RSS was 2.196 GiB. Available disk stayed above 69 GiB.
The final package configuration therefore uses concurrency two. The earlier concurrency-one measurement remains the serial reference; each locale/surface still gets a separate file/process and the same cache cleanup and assertions.
## Golden compare/capture time budget
At head `45a4e32`, [Golden run 38077035195](https://github.com/coco-de/demo-ecommerce-grocery/actions/runs/38077035195) reached the forty-minute job budget with compare still running. Its compare step started at 18:46:26 UTC and was cancelled at 19:25:19 UTC; job preparation began at 18:45:06 UTC. Six package compares completed and order-console was still in progress, leaving four further packages. A complete compare or difference classification is not claimed.
The same-head [Pages run 38077035136](https://github.com/coco-de/demo-ecommerce-grocery/actions/runs/38077035136) completed a normal Linux capture of all eleven packages from 18:48:28.1178029 UTC to 19:15:32.6746240 UTC: 27 minutes 4.56 seconds. Its pre-deployment smoke passed eight checks.
Compare's completed package timings and its remaining work, followed by the measured capture and gallery/artifact steps, indicate roughly 85 minutes for a full failing compare plus review capture. Only the Golden compare job budget is therefore changed to 100 minutes, including setup margin. The update job remains thirty minutes. The always-run capture of the expired job was force-cancelled rather than counted as compare proof.
The completed timeout/concurrency changes must be pushed before the next full Linux comparison can produce the exhaustive classification CSV. Renderer sources, assets, supplier pins and baseline PNGs are unchanged by this timing adjustment.
## Complete Linux classification and update budget
At source head `141c27ac72873ce0010029425fd5811db108ed0d`, the retained
[Golden run 38080025704](https://github.com/coco-de/demo-ecommerce-grocery/actions/runs/38080025704)
finished its full eleven-package compare in 47m42s. It passed 116 variants and
failed 6,676: 6,168 missing new baselines and 508 existing pixel mismatches.
Normal capture/gallery then passed all 6,792 variants in 23m38s, with no render
errors, duplicate paths, missing images or retired baseline paths.
[s4-golden-classification.csv](s4-golden-classification.csv) records all 6,676
changed paths, causes, owning W4 work, source evidence, source/run identifiers,
old/capture SHA-256, dimensions and changed-pixel counts/bounds. The changed
path set equals the complete compare log's failure path set. The 624 existing
baselines comprise 508 changed and 116 byte-identical images; all dimensions
are unchanged. The surviving 220 comparator master/test pairs are each
byte-identical to their normal capture. Flutter overwrites variant-only diff
filenames across scenarios, so those 220 pairs are not the exhaustive count.
| Classification | Paths | Owner |
|---|---:|---|
| Additional language variants of existing scenarios | 1,200 | W4-08 |
| Additional state/permission/long-input scenarios | 4,968 | W4-08 |
| Existing localized date/time display | 100 | W4-07 |
| Existing dispatch date/time/currency display | 24 | W4-07 |
| Existing feature-owned B projection plus money/date formats | 336 | W4-06 / W4-07 |
| Existing review/product display projection | 48 | W4-06 |
The existing kernel 24, produce-browse 24, error/empty 48 and inbox 20 variants
remain byte-identical. W4-10 did not change the palette, and the shared font
harness is not assigned as the cause of every existing mismatch. Representative
Linux date/currency/name/address changes were reviewed against their source;
Arabic option/stock, Japanese cutoff wrapping and the valid 300-character
English review fixture were also opened. Complete machine coverage and these
representative visual checks are reported separately.
Only the update job timeout is changed in this checkpoint, from 30 to 60
minutes. It renders the same eleven-package matrix; the measured 23m38s capture
and prior 27m04s Pages capture leave little margin at 30m for preparation,
baseline writes and bot commit/push. Sixty minutes provides roughly twice the
measured capture budget. Renderer/fixture sources, assets, supplier baseline
v1 and generated source are unchanged. The coordinator-approved next step is one Linux
`[update-goldens]` update of exactly the classified paths, followed by bot-path
and normal-capture pixel verification and same-final-head CI/compare/Pages.
## Linux bot update verification
The single authorized [Linux update run 38085201259](https://github.com/coco-de/demo-ecommerce-grocery/actions/runs/38085201259)
succeeded at parent `4baaeeee1e877e4d21ab7f79ae2baedffa814fba` and wrote bot
commit `f917d2431f220231ae456ee8bc6e5631215eb357`. Its 6,676 changed paths are
all baseline PNGs and exactly equal the classification CSV; no extra path,
retired baseline or unclassified change exists. All 6,792 committed baselines
match the prior normal Linux capture byte for byte, including dimensions and
RGBA pixels: byte/dimension/pixel mismatches are each zero.
The update job took32m22s (update step29m42s), beyond the former30m job budget;
it completed under the approved60m limit.
Renderer, fixture, fonts/assets, supplier pins, generated source and migration
history remain unchanged from the verified Linux capture input. Only the
classification documents, update timeout and baseline PNGs differ. The full
per-path hash/dimension proof is retained in PC-A's dispatch evidence directory
`/Users/dongwoo/orca/toolchains/637/xhost/s4-659-merge-train/dispatch-ctx_36ac5850b751/`
(`linux-update-verification.json`, `render-inputs-unchanged.json`).
Bot-head CI38087245166, Golden38087245245 and Pages38087245267 returned
`action_required` before running. They are superseded without workflow approval
or failed-job rerun. This authorized writer commit records the completed
path/pixel review and triggers fresh CI, Golden compare and Pages at one final
head; merge remains conditional on those same-head results all being green.