feat(browse): add record command for video evidence of interactive bug repros

[itstimwhite]

Why

Static screenshots can't capture the timing of a UI bug. Form-submission order, async loading flicker, drag/drop, scroll-triggered behavior, focus management, dialog timing — these all benefit from a few seconds of video. Annotated screenshots stay the right answer for static bugs (typo, clipped UI, missing element), but interactive bugs deserve a .webm so the developer fixing them can see the exact repro shape.

agent-browser (Vercel Labs) ships a dogfood skill that records video evidence as part of its structured QA workflow. I wanted the same evidence shape available behind $B — including from inside /qa and /qa-only — so this PR adds the primitive at the browse layer and leaves the QA-side flag for a follow-up.

What

A new record meta-command with three subactions:

$B record start [path] [--size WxH]    # rebuild context with recordVideo
$B record stop                          # close context, flush .webm, return paths
$B record status                        # is recording active?

Wraps Playwright's BrowserContext.recordVideo option. Implementation hooks into the existing recreateContext() save/restore path, so cookies, URLs, and open pages survive both start and stop. @e refs are invalidated by the recreate (same caveat as viewport --scale); record stop prints a hint reminding you to re-snapshot.

Defaults output to a timestamped subdirectory under TEMP_DIR so concurrent recordings never collide. Custom paths pass through the standard validateOutputPath policy.

Single-recording invariant: calling record start while already recording auto-stops the prior recording. Headed mode rejects with a clear error (the user can use their OS screen recorder in headed sessions).

Tests

9 integration tests in browse/test/record.test.ts:

• record status reports not-recording before start
• record stop is a no-op when not recording
• record start → activity → record stop produces a non-empty .webm (verified by reading the EBML magic bytes 0x1A 0x45 0xDF 0xA3)
• Browser remains functional after stop (state preserved across the context recreate)
• record start while already recording auto-stops the prior recording
• Unknown flag, malformed --size, missing action, unknown subaction all reject with usage errors

All passing locally:

 9 pass
 0 fail
Ran 9 tests across 1 file. [2.29s]

Existing browse/test/commands.test.ts (223 tests) still green.

The pre-existing snapshot.test.ts flake (closetab last tab auto-creates new) reproduces on main without these changes; not related.

Commits (bisect-friendly)

1. feat(browse): add record command … — implementation only (browser-manager.ts, commands.ts, meta-commands.ts, new record.test.ts). 352 insertions.
2. docs(browse): document record command in SKILL.md.tmpl + regenerate — SKILL.md.tmpl change + regenerated browse/SKILL.md, top-level SKILL.md, gstack/llms.txt. 67 insertions.

Template change and generated-doc regeneration intentionally split, per the gstack contributor guidance.

Out of scope

• VERSION bump and CHANGELOG entry — left for the merge so you can write the entry in your voice. The implementation includes a NEW_IN_VERSION: 'record': '1.35.0.0' so the unknown-command hint works the moment this lands; adjust if you cut a different version.
• /qa --evidence-per-finding flag (the QA-side complement) — separate PR coming, since it's a template-only change to two SKILL.md.tmpl files and is easier to review on its own.
• Auto-cleanup of recorded .webm files. They're evidence; outliving the session is the point.