Per-suite flake sentinels: specs + preview e2e

Status summary: Extends the flake pipeline's self-proof to the suites most likely to actually flake. Each suite gets its own deliberately ~10%-flaky, month-gated sentinel (distinct test names → distinct dashboard rows, so a suite whose sentinel reads 0% has broken plumbing), plus the two-line CI wiring that makes its records flow: FLAKE_RECORD_DIR on the lane + a flake-records-<suite> artifact upload. Ingestion is already suite-agnostic (#2582) — no server changes.

Checklist#

specs/flake-sentinel.spec.ts — playwright sentinel (verified through the real spec runner: expected-fail registration accepted, record line written; cosmetic wart: playwright shows the wrapper's file as the test location)
apps/os/e2e/vitest/flake-sentinel.e2e.test.ts — preview e2e sentinel (verified under the e2e vitest config; it flaked on the verification run and stayed green, recording the flake-fail)
preview lanes: FLAKE_RECORD_DIR per lane in scripts/preview/preview.ts (playwright → test-results/flake-records/specs, vitest e2e → .../preview-e2e), guard tests updated
cloudflare-previews.yml: flake-records-specs + flake-records-preview-e2e artifact uploads (if: always(), if-no-files-found: ignore, overwrite: true)
docs/testing.md: note the per-suite sentinel convention + the stance that createFlake replaces retries only for tests that opted in (unwrapped specs keep playwright retries)
createFlake pins per-test retry: 0 on vitest (surfaced by the first real preview run: the e2e suite's retry: {count:1, delay:5000} re-ran the sentinel after the wrapper's green throw — vitest retry fires before the .fails inversion — recording every green outcome twice; fixture child now runs with suite retry to pin it)
ingestion: drop the Depot run-status filter (also surfaced live: Depot's per-sha run status settles only after the LAST check completes, and that check's webhook beats the flip, so status: ["finished","failed"] in ListRuns permanently skipped the last-completing check's artifacts — always the slow preview check, i.e. exactly these two new suites)
first real adoptee: specs/mobile/chat-photos.spec.ts wrapped with createFlake (defeated its playwright retry twice on 2026-09-03 — Timeout …ms exceeded waiting for getByLabel('phone-screenshot.png'); pattern scoped to that signature, 120s wrapper deadline under the 240s spec timeout)

Post-merge#

Confirm both new rows appear on the Flake dashboard issue after a preview run on main

Implementation log#

Preview slot 17 was damaged (missing container classes, upstream Cloudflare gap); repaired by deleting the os-preview-17 worker via the Cloudflare API (no queue consumers were attached) and letting the bootstrap recreate it on the next deploy. No dedicated repair script exists — erase-data deliberately leaves workers in place.
Run rs7dwp1l27 attempt 3 proved the wiring end to end: both sentinels ran, recorded, and uploaded (flake-records-specs 276B, flake-records-preview-e2e 286B); the job's only real failure was an unrelated mobile chat-photos flake.
Those two artifacts were then never ingested — replaying the deployed ingestion steps locally (ListRuns/ListArtifacts/download/unzip/parse) all succeeded, which isolated the run-status race fixed above.