ci: shard the browser e2e suites across runners

[Rinat S (rsolmano)]

Problem

CI takes ~22 min on every PR and merge-queue entry (measured over the last successful runs, e.g. run 37349114307). Almost all of it is Playwright execution in three jobs, which run the same ~464 no-agent specs against three hosts:

Check	Wall time	Of which test step
Desktop smoke & e2e	22.3 min	1215 s, 1 worker
Binary smoke & e2e	20.0 min	1126 s, 1 worker
E2E (no-agent)	13.6 min	755 s, 2 lanes (4-vCPU runner)

Setup (checkout, install, Playwright, builds) is only ~1–2 min per job.

Approach

Split each browser suite across several runners and keep the ruleset's required checks stable through aggregate gate jobs.

• Source suite: 2 jobs. Each runs one slice via THINKRAIL_E2E_JOB_SHARD=k/N, and its local lanes subdivide that slice into global Playwright shards (k-1)·L+i of N·L, so the jobs together cover exactly the whole suite.
• Binary and desktop suites: 4 jobs each, using Playwright's --shard=k/4 (the runners already forward argv). The artifact configs now distribute individual tests (fullyParallel: true, still workers: 1), so the split is even (116 × 4) instead of file-sized.
• Gates: the ruleset requires E2E (no-agent), Binary smoke & e2e and Desktop smoke & e2e. Those names now belong to small gate jobs (needs: the matrix, if: !cancelled()) that pass only when every shard succeeded, so no ruleset change is needed.

Trade-off: wall time drops from ~22 to ~7–8 min (estimated from the step timings above), while total runner time grows ~20% (~62 → ~74 min per run), because each shard repeats the ~1–1.5 min setup. Running the full artifact suites only in the merge queue was considered and left out: with sharding it no longer shortens PR wall time; it would only save runner minutes.

Changes

• .github/workflows/ci.yml: e2e-shard (2), binary-shard (4) and desktop-shard (4) matrices with fail-fast: false and per-shard report artifact names, plus the three gate jobs under the existing required names.
• e2e/shardPlan.ts, e2e/run.ts: parseJobShard / laneShardArgs. The runner applies the job slice in both serial and lane modes, and rejects THINKRAIL_E2E_JOB_SHARD combined with an explicit --shard.
• playwright.config.ts: fullyParallel is also on when a job slice is set.
• e2e/artifactPlaywright.ts: fullyParallel: true for binary and desktop. No spec uses beforeAll/afterAll/describe.serial, and the source lanes already run the same specs split per test.
• e2e/SPEC.md, .github/SPEC.md: job-slice contract, per-test artifact sharding, and the gate-name ↔ ruleset coupling.

Screenshots

Not applicable — no frontend changes.

Related issues

None.

Checklist

• Fast gates pass: bun run lint, bun run typecheck, bun run test
• E2E suite passes for app-affecting changes (bun run e2e, or bun run e2e:full when touching agent behavior) — not app-affecting (test-harness + CI only); targeted sharded runs below, and this PR's own CI runs every shard
• Before/after screenshots are included for frontend changes, or marked not applicable
• Relevant SPEC.md / top-level specs updated to reflect any boundary, contract, or behavior change
• I have read the Contributing guide and agree to the Code of Conduct

Testing

• bun run lint → clean
• bun run typecheck → 18/18 successful
• bun run test → 18/18 tasks, 0 fail
• bun run check:deps / check:boundaries / check:seams / check:spec-surface → OK
• playwright test e2e/shard-runner.spec.ts → 8 passed (incl. new job-slice test)
• bun run e2e --list → 470; THINKRAIL_E2E_JOB_SHARD=1/2 and 2/2 → 235 each; combined with --shard=1/2 → rejected (exit 1)
• playwright test -c playwright.{binary,desktop}.config.ts --list --shard=k/4 → 116 each of 464
• bun run build:binary && bun run e2e:binary --shard=1/4 → 116 passed (4.8 min)
• THINKRAIL_E2E_JOB_SHARD=2/2 bun run e2e → 232 passed (8 local lanes → global shards 9–16 of 16)
• @action-validator/cli .github/workflows/ci.yml → valid
• Desktop shard not run locally (needs Linux + Xvfb); this PR's CI exercises all four.

[jetbrains-air]

Approved — ready to merge.