fix: use beginFrame for headless recording capture

[carlesandres]

Summary

This PR narrows the core headless recording fix from vercel-labs#12.

Credit to @sld0Ant for the earlier investigation and implementation work that identified the correct fix direction. This PR keeps only the begin-frame-control capture-path fix, the aligned PNG source-frame update, and focused test coverage so it is easier to review independently.

If you have time, I’d also welcome feedback from @sld0Ant on whether this captures the intended minimal version of the earlier fix.

Problem

webreel already launches headless Chrome with --enable-begin-frame-control, but the recorder still captures frames with:

• Page.captureScreenshot()
• JPEG q60 source frames

That creates two problems:

1. Page.captureScreenshot() is the wrong capture API for begin-frame-control mode
2. JPEG source frames introduce visible artifacting before the final video encode

Changes

This PR makes the narrowest recording-path fix I could extract from vercel-labs#12:

• add HeadlessExperimental to the CDP client type
• call HeadlessExperimental.enable() during recording setup
• add a pre-record frame pump while the page is loading in begin-frame-control mode
• stop that pre-record frame pump both before Recorder.start() and during failure cleanup
• replace Page.captureScreenshot() with HeadlessExperimental.beginFrame()
• switch source capture from JPEG to PNG with optimizeForSpeed
• save debug frames as .png

Why this helps

This fixes the headless capture path for begin-frame-control mode and improves source-frame quality for final exports.

The scope here is intentionally narrow. This PR is not claiming full 4k support or universal high-resolution throughput. It is only claiming:

• the correct headless capture API for begin-frame-control mode
• cleaner PNG source frames instead of JPEG q60
• reduced JPEG artifacting in final exports

Tests

Added focused coverage for:

• recorder using HeadlessExperimental.beginFrame() with PNG screenshots
• debug frames being written as .png
• HeadlessExperimental.enable() happening before Recorder.start()
• pre-record frame pump cleanup when setup fails before recording starts

The small vitest.config.ts alias change is test-only plumbing for that runner coverage. It lets Vitest resolve the workspace @webreel/core package from source instead of requiring a built dist output first.

Verification run locally:

• pnpm --filter @webreel/core test
• pnpm --filter webreel test
• pnpm type-check

Docs

Updated the user-facing --frames wording to match current behavior:

• CLI help now says raw frames are saved as PNGs
• README and skill docs now describe .webreel/frames/ as PNG source frames in headless recording mode

Non-goals

This PR intentionally does not include:

• compositor queue changes
• GIF pipeline rewrites
• benchmark scripts
• broader hi-res throughput claims
• unrelated refactors from fix: deterministic headless recording, lossless capture, and compositor optimizations vercel-labs/webreel#12