ci: extract shared Playwright-server harness into a composite action (#1752)

[johnrtipton]

Closes #1752.

Problem

nav-hooks-guard and playwright-tests duplicated ~9 setup steps (Python/uv/cache/uv sync/maturin build/Playwright install/migrate/start server/wait). #1753 had to patch the missing maturin develop step into nav-hooks-guard specifically because the copy had drifted out of sync — textbook parallel-path drift (#1646).

Fix (item 3)

Extract the shared harness into a composite action .github/actions/djust-playwright-server. Both jobs now actions/checkout → uses: ./.github/actions/djust-playwright-server → run their own test step → upload server.log on failure → stop server.

• One definition can't drift a step out of one job.
• Jobs keep their identity: playwright-tests stays continue-on-error: true (optional), nav-hooks-guard stays blocking; test-summary needs/AND-condition unchanged.
• Net −31 lines (+95/−126).

Items 1 & 2

• Item 1 (maturin build) shipped in ci(test): build Rust extension in nav-hooks-guard before running the guard (#1752) #1753 — now lives in the composite, so it can't drift again.
• Item 2 (blocking-vs-soak) resolved by decision: keep nav-hooks-guard blocking. The soak concern (Free-threaded hardening / perf polish (follow-up to #1432) #1534) is retroactively satisfied — it has run green across Refs #1745 -- Harden nav/hooks regression guard #1748, ci(test): build Rust extension in nav-hooks-guard before running the guard (#1752) #1753, and fix(render): close-side </div> tolerance + consolidate dj-root scanners (#1751) #1754, and the condition-based waits (Refs #1745 -- Harden nav/hooks regression guard #1748) reduce flake.

Verification

• Local: both YAML files parse; every composite run step declares shell: bash; both jobs preserve checkout→composite ordering + their stop/upload steps + continue-on-error settings.
• The empirical proof is this PR's own CI: both playwright-tests and nav-hooks-guard must run green through the composite (a new runner-only path; per Free-threaded hardening / perf polish (follow-up to #1432) #1534, ≥1 runner iteration may be expected for composite-action quirks).

🤖 Generated with Claude Code

[johnrtipton]

Code Review (pipeline-run Stage 7) — APPROVE

CI-only refactor; closes #1752. No 🔴.

What it does: extracts the 9 shared setup steps into .github/actions/djust-playwright-server/action.yml (composite); playwright-tests and nav-hooks-guard become checkout → uses: ./.github/actions/djust-playwright-server → <own test> → upload-log-on-failure → stop-server.

Correctness:

• Empirical canary (feat: improve error messages for common LiveView mistakes #252): this PR's CI ran both jobs green through the composite (nav-hooks-guard 3m30s, playwright-tests 3m30s) — proves the background server started in a composite step persists into the caller's test step, and the local-action path resolves after checkout.
• Every composite run step declares shell: bash (required for composite run steps); uses: steps (setup-python/uv/cache) don't need it. Verified.
• Job identities preserved: playwright-tests keeps continue-on-error: true (optional), nav-hooks-guard has no continue-on-error (blocking); test-summary needs/AND-condition reference the same job names → gating unchanged.

Concern / edge cases considered:

• Cache key unified to ${{ runner.os }}-python-playwright-… (was a separate -nav-hooks- key). Both jobs now share one cache — harmless (same deps), and restore-keys still falls back to -python-.
• The historical demo-checks decoupling note (ci: dogfood djust_check against the demo project (catch dead-button / deprecated-attr drift) #1708/ci: promote demo djust_check dogfood from continue-on-error to a blocking gate #1713) was preserved inline in playwright-tests.
• The composite's maturin build (item 1) now lives in ONE place → can't drift out of one job again (the fix(auth): wrap per-event check_object_permission in sync_to_async (#1638) #1646 failure mode that caused ci(test): build Rust extension in nav-hooks-guard before running the guard (#1752) #1753).

Item 2 (blocking-soak): resolved by decision — keep nav-hooks-guard blocking; green across #1748/#1753/#1754 + #1755 satisfies the #1534 soak.

APPROVE

[johnrtipton]

Retrospective — PR #1755 (#1752)

Task: extract the shared Playwright-server harness into a composite action (#1752 item 3); resolve item 2 (keep blocking) by decision. Item 1 (maturin) shipped in #1753.

Quality: 5/5

Fixes the root cause of the whole #1752 thread — the duplication that let the maturin step drift out of one job — not just its symptom.

What went well

• Root-cause over symptom. ci(test): build Rust extension in nav-hooks-guard before running the guard (#1752) #1753 patched the missing maturin step into nav-hooks-guard; this PR removes the duplication that allowed it to go missing. The composite is one definition; a step can no longer exist in one copy and not the other (fix(auth): wrap per-event check_object_permission in sync_to_async (#1638) #1646 parallel-path-drift, now applied to CI not just code).
• Empirical canary (feat: improve error messages for common LiveView mistakes #252) for free. A composite action has real first-use unknowns (does a backgrounded server survive the composite→caller step boundary? does the local-action path resolve post-checkout?). This PR's own CI answered both: nav-hooks-guard and playwright-tests both ran green through the composite. No guessing.
• Identity preserved through the refactor. continue-on-error (playwright optional / nav blocking) and the test-summary AND-condition were left untouched because the jobs stayed jobs — the lower-risk composite-action choice over a reusable workflow (which would have restructured the gating).
• Net −31 lines while adding a reusable building block.

What could be improved

• The duplication existed for several PRs (tech-debt: dogfood dj-navigate cross-view flow + a client-hook (3rd-party lib) in the demo so demo-checks catches nav/hydration/hook regressions #1742 → Refs #1745 -- Harden nav/hooks regression guard #1748 → ci(test): build Rust extension in nav-hooks-guard before running the guard (#1752) #1753) before being factored out; the drift (ci(test): build Rust extension in nav-hooks-guard before running the guard (#1752) #1753) is what finally motivated it. Lesson reinforced: when a CI job is added by copy-pasting an existing one (Refs #1745 -- Harden nav/hooks regression guard #1748 copied playwright-tests), factor the shared part then, not after a step drifts.

Verified

• Both YAML files parse; composite run-steps all declare shell: bash; both jobs preserve checkout→composite ordering + stop/upload + continue-on-error. CI fully green — both Playwright jobs through the composite, plus rust/python/demo-checks.

RETRO_COMPLETE