fix(ci): restore documented CI job graph in test.yml (#51)

* fix(ci): restore documented CI job graph in test.yml

build and playwright-tests both drifted to gating on rust/vitest-tests
in addition to deno-lint, producing a serial rust -> playwright ->
build chain that docs/builds.md never described. Neither job consumes
artifacts from rust or vitest-tests, so narrow both back to what the
docs already document: deno-lint alone gates build, and deno-lint plus
the frontend path filter gates playwright-tests.

Also updates the dependency-graph doc itself, which still said "five
jobs" and listed playwright-tests as fully independent -- stale since
zig, shadow-diff, and changes were added and playwright-tests already
depended on deno-lint.

* chore(backlog): close out TASK-351 with real PR-run CI timing evidence

Author: pythoninthegrass

.github/workflows/test.yml

@@ -154,9 +154,10 @@ jobs:
 
   build:
     name: Build (${{ matrix.platform }})
-    # Gate on lint/format/test jobs so builds don't burn CI minutes
-    # when basic checks fail.
-    needs: [deno-lint, rust, vitest-tests, playwright-tests]
+    # Gate on deno-lint only -- rust/vitest-tests/playwright-tests run in
+    # parallel and don't produce artifacts the build matrix consumes, so
+    # serializing on them just adds wall-clock (see docs/builds.md#test-workflow-dependency-graph).
+    needs: [deno-lint]
     if: ${{ !failure() && !cancelled() }}
     strategy:
       fail-fast: false
@@ -278,7 +279,9 @@ jobs:
 
   playwright-tests:
     name: Playwright E2E Tests
-    needs: [changes, deno-lint, rust, vitest-tests]
+    # Runs against a Vite preview build with mocked IPC -- no Tauri/Rust
+    # artifact dependency, so it only needs deno-lint plus the change filter.
+    needs: [changes, deno-lint]
     if: ${{ !failure() && !cancelled() && (needs.changes.outputs.frontend == 'true' || github.event_name == 'workflow_dispatch') }}
     runs-on: [macOS, ARM64]
     timeout-minutes: 20

backlog/tasks/task-351 - Restore-the-documented-CI-job-graph-in-test.yml.md

@@ -1,9 +1,10 @@
 ---
 id: TASK-351
 title: Restore the documented CI job graph in test.yml
-status: To Do
+status: Done
 assignee: []
 created_date: '2026-09-11 00:38'
+updated_date: '2026-09-14 07:00'
 labels: []
 dependencies: []
 type: bug
@@ -18,9 +19,27 @@ Cross-platform build verification should start as soon as deno-lint finishes rat
 
 ## Acceptance Criteria
 <!-- AC:BEGIN -->
-- [ ] #1 build job gates only on deno-lint
-- [ ] #2 playwright-tests job gates only on changes and deno-lint
-- [ ] #3 A PR run shows the build matrix starting within ~3 minutes of deno-lint completing
-- [ ] #4 Total CI wall-clock recorded before and after the change in the task notes
-- [ ] #5 docs/builds.md dependency graph re-verified as accurate against the workflow file
+- [x] #1 build job gates only on deno-lint
+- [x] #2 playwright-tests job gates only on changes and deno-lint
+- [x] #3 A PR run shows the build matrix starting within ~3 minutes of deno-lint completing
+- [x] #4 Total CI wall-clock recorded before and after the change in the task notes
+- [x] #5 docs/builds.md dependency graph re-verified as accurate against the workflow file
 <!-- AC:END -->
+
+## Implementation Notes
+
+<!-- SECTION:NOTES:BEGIN -->
+## CI wall-clock: before/after
+
+**Before (drifted graph)** — most recent real PR run on the old graph, [run 34811970745](https://github.com/pythoninthegrasses/mt/actions/runs/34811970745) (TASK-355.6's PR): `rust` failed on a pre-existing, unrelated bug (`resource path binaries/mt-zig-core-aarch64-apple-darwin doesn't exist`). Because `build` and `playwright-tests` both depended on `rust`, they were **skipped entirely** rather than delayed — total run wall-clock 1m50s, but zero build/E2E signal produced for the PR. That's the sharpest cost of the drift: one unrelated job failure silenced build+E2E for the whole PR, not just slowed them.
+
+For a wall-clock (not skip-cascade) comparison, the last fully-successful run on the old graph, [run 24358563505](https://github.com/pythoninthegrasses/mt/actions/runs/24358563505) (2026-04-13, predates the `zig`/`shadow-diff`/`changes` jobs): total 17m8s. `build(macos)` didn't start until 18:02:44 — 2m18s after `deno-lint` finished (17:58:26) — because it waited on `playwright-tests` (finished 18:02:42), which itself waited on `rust`/`vitest-tests`. In that run rust/vitest/playwright happened to be fast (~1-2 min each), so the observed serial penalty was minutes, not the worst-case ~40 min the task description assumes for slower runs — but the mechanism (build gated on jobs it doesn't need) is the same regardless of their duration.
+
+**After (this fix)** — [PR #51's own run, 34815076214](https://github.com/pythoninthegrasses/mt/actions/runs/34815076214): `deno-lint` finished 06:49:56. `build(linux)` started 06:50:10 (+14s), `build(windows)` started 06:50:26 (+30s), `build(macos)` started 06:51:22 (+1m26s) — all well within the ~3min target (AC#3), and critically, **all three started even though `rust`/`vitest-tests`/`shadow-diff` failed** (same pre-existing zig staging bug as the before-case), because `build` no longer depends on them. Total run wall-clock 9m34s — longer in absolute terms than the before-cascade-skip case specifically because build now actually runs to completion three times instead of being dropped; that's the intended behavior, not a regression.
+
+`playwright-tests` was correctly skipped in the after-run (completed 06:49:56, essentially instantly) because this PR touches no `app/frontend/**` paths, not because of any upstream failure — confirms AC#2's `changes` gating works independently of `rust`/`vitest-tests` now.
+
+## AC#5 — docs/builds.md re-verification
+
+Updated `docs/builds.md`'s "Test Workflow Dependency Graph" section in the same commit: it said "five jobs" and listed `playwright-tests` as fully independent, both stale (workflow now has eight jobs — `zig`/`shadow-diff`/`changes` were added since the doc was last touched — and `playwright-tests` already depended on `deno-lint` even before this fix). Section now names all eight jobs and correctly shows `build` and `playwright-tests` both gated on `deno-lint` only (plus `playwright-tests` on `changes`), matching `test.yml` exactly.
+<!-- SECTION:NOTES:END -->

docs/builds.md

@@ -612,17 +612,20 @@ Runner assignment is optimized for developer iteration speed (PR/push), not rele
 
 #### Test Workflow Dependency Graph
 
-The test workflow (`test.yml`) runs five jobs. The build matrix only gates on `deno-lint` — the test jobs run in parallel and do not block cross-platform build verification:
+The test workflow (`test.yml`) runs eight jobs. The build matrix and Playwright both gate only on `deno-lint` — the other test jobs run in parallel and do not block cross-platform build verification:
 
 ```text
 deno-lint ──► build(macos, linux, windows)
+deno-lint ──► playwright-tests (also gated on `changes`: frontend-paths filter)
 
-rust             (independent)
-vitest-tests     (independent)
-playwright-tests (independent)
+rust         (independent)
+vitest-tests (independent)
+zig          (independent)
+shadow-diff  (independent, continue-on-error)
+changes      (independent — feeds playwright-tests' path filter only)
 ```
 
-The `rust`, `vitest-tests`, and `playwright-tests` jobs do not produce artifacts consumed by the build matrix (`cargo check` per platform). Decoupling them from the build reduces wall-clock time by allowing cross-platform checks to start as soon as `deno-lint` completes (~2-3 min) rather than waiting for the slowest test job (~15-20 min).
+The `rust` and `vitest-tests` jobs do not produce artifacts consumed by the build matrix (`cargo check` per platform), and `playwright-tests` runs against a Vite preview build with mocked IPC rather than any Tauri/Rust artifact — so none of them need to gate `build` or each other. Decoupling them reduces wall-clock time by allowing cross-platform checks and Playwright to start as soon as `deno-lint` completes (~2-3 min) rather than waiting for the slowest test job (~15-20 min). `zig` and `shadow-diff` are deliberately absent from every other job's `needs:` so Zig-core work never extends the critical path (see the inline comments in `test.yml` for each job's rationale).
 
 #### CI Setup Modes