ci: shard Playwright E2E by measured suite timings (follow-up to #383) #434

Problem

After #383 parallelised Go tests, E2E and the image check, the Playwright E2E job is now the critical path: about 25 min of the 25–28 min PR runs. With several PRs in flight, jobs also queue behind each other, so total wall time often reaches 40–50 min.

#383 deliberately deferred E2E sharding until we had measured timings ("sharding should follow measured timings from a live run rather than copying upstream's assignments").

Proposal

Split the E2E job into N shards (start with 3) that run in parallel, based on measured per-suite durations from recent green runs:

• Extract per-suite wall time from the last several green 🎭 Playwright E2E Tests logs (the === E2E SUITE: … === markers), and put the table in the PR.
• Assign suites to shards to balance total time (greedy longest-first). Keep the assignment in a checked-in manifest, so it is reviewable and updatable.
• Each shard prepares the same fixture as today (freshen, seed SQL, corescope-migrate, seeds) and runs its own local server on its own port.
• Fail-closed: the existing 🏗️ Build & Publish Docker Image gate must require every shard to succeed. Skipped, cancelled or empty counts as failure, as ci: parallelize independent checks with fail-closed gate #383 established.
• A guard test fails CI if a new E2E suite file is not assigned to any shard, so suites cannot silently drop out.
• Keep the fork guards (9 in deploy.yml, 1 in release-fast-path.yml). No publish or deploy from the fork.

Acceptance

• The per-suite timing table and the shard assignment are in the PR.
• A PR-triggered run shows every shard green, and the measured E2E wall time is down versus the current ~25 min. Report before/after for at least two runs.
• The unassigned-suite guard and the fail-closed gate are tested (mutants shown red).
• Consider the runner-minute cost: shards repeat the setup. Report total runner minutes before and after.