feat(record): per-test retries + scene-failure hint + slow-scene warning

[shreyaskarnik]

Summary

First half of the long-demo flake mitigation roadmap. Pure additive — no behavior change unless callers opt in.

• video.retries: N config + --retries N CLI flag — embeds Playwright's built-in retries in the generated playwright.record.config.mjs so a transient failure (browser hiccup, race in showOverlay, intermittent network) gets a free retry without manual intervention.
• Failing scene hint — when the recording subprocess fails, the error message appends the last scene name from .scene-progress.jsonl so users can map the failure to a scene without reading the Playwright stack trace.
• Slow-scene warning — after a successful pipeline, scenes whose duration is more than 1.75× the median are surfaced as a console warning. Useful intermittent-flake signal — a scene that drifts from 6s → 18s on retries is the most likely culprit.

Scope

This is v1 of the long-demo flake roadmap. v2 (scene-isolated recording with state checkpoints + --resume-from) is deliberately deferred — it's ~5× the scope and the design will be better informed by what failure modes v1's diagnostics surface in production. v1 alone catches a lot of transient failures via the Playwright retry plumbing.

Test plan

• npm test — 639 tests pass (3 new tests covering retries embedding: default 0, explicit value, clamp/floor of fractional values)
• Local npx argo pipeline ci-smoke --retries 1 runs cleanly; slow-scene warning fires correctly on the trailing-wait scene
• No behavior change for callers that don't set retries (default remains 0, current behavior)
• CI matrix passes (chromium / firefox / webkit) — should be untouched, but verifying