revert: stop releasing a WebGL context when its clip's window has ended - #7020

Requesting changes at 095217fd. The code is fine; this is a merge-gate block.

Blockers

1. The PR is held, but approving it would merge it. The body says it is "held, not for merge, until a repro shows 3D blank on 0.9.x and back with the revert". In this lane, one approval on the head is enough for the merge loop to merge. The PR is not a draft and has no hold label. Its own evidence so far points the other way: two Three.js fixtures, with 1 and 3 workers, rendered byte-identical frames on 0.8.230 and 0.9.2. Approving now would ship the revert without the evidence the PR asks for.
2. There is no fails without the fix: line. The PR is a fix for the "3D vanishes from exports" report, so it has to name a test that fails without it. That test is the missing repro: a Three.js/WebGL fixture that renders blank on 0.9.x and renders 3D with this revert, for example with several workers or a nested clip whose window ends early.

What I checked (no code concerns)

• The diff is the exact inverse of #6984 (26d6e3ee), apart from the packed size budget (461788 → 456397). It is current with main (ahead 1, behind 0).

To clear this: add the repro test and the fails without the fix: line, then re-pin the head here. I'll re-review it. Until then, keep it in draft or add a hold label so another approval can't merge it.

Approving at ebfb8fba. I found no blockers. The earlier hold is lifted by the maintainer's decision to ship it today.

• Change: it removes only the three places where frameCapture.ts wires in the release (the import, evaluateOnNewDocument(recordWebglContexts) and releaseEndedWebglContexts(page, quantizedTime)). webglRelease.ts, its unit tests, clipWindow and the size budget stay for the follow-up fix. Nothing outside tests still calls the module.
• Fail-without-fix check: the fails without the fix: line names one test, and I ran it both ways.
  • With this head, webglRelease.test.ts passes (4/4).
  • With frameCapture.ts put back as on main, the named test fails with expected [ '3d' ] to deeply equal [], meaning loseContext was called once, as the body says.
• Dead code: webglRelease.ts is now only imported by its test. I ran fallow audit --base origin/main --fail-on-issues locally on this head: "No issues in 50 changed files".
• Main: the branch is one commit behind main, and that commit is #7015 (CLI catalog files only), so there's no overlap.
• Nit (not blocking): the test's page mock only acts on functions whose source contains __hf_gl_contexts. That makes it a guard against the wiring coming back, not a 3D render repro. The repro is still owed for the real fix.

CI: CI, regression, Windows render verification and CodeQL were still pending or running when I reviewed. The runs that had finished passed, apart from one cancelled duplicate PR captures run (its rerun passed).