browser_probe video visibility discovery is O(videos × duration) and hurts GCP Cloud Run plan time

[valeriangalliat]

Describe the bug

When compile injects data-hf-auto-start (videos with no authored data-start), discoverVideoVisibilityFromTimeline walks the GSAP timeline once per video. Coarse seeks ≈ V × D / 0.1.

This mainly shows up on GCP Cloud Run (CPU / SwiftShader). Local Metal finishes the same loop quickly. On Cloud Run, plan stage wall time grows with video count. Heavy real compositions can push the probe past the ~5 min Puppeteer protocolTimeout.

Optional patch (same windows, one seek per step): #3233

Link to reproduction

https://github.com/ArcadeHQ/hyperframes-repros/tree/patch/visibility-scrub

Steps to reproduce

Local (path check only):

1. Clone the branch. Pin unpatched hyperframes@0.8.3 / 0.8.4.
2. npm run render -- --debug -f 1 --quality draft --output renders/unpatched.mp4
3. Confirm:

   Launching browser for composition probe... { reasons: [ 'auto-start video(s)' ] }
   Discovering video visibility windows... { videoCount: 5 }
   

Do not add data-start. lint media_missing_data_start is expected.

Cloud Run (where the cost matters):

1. Deploy unpatched HyperFrames with hyperframes cloudrun deploy (or an equivalent unpatched image).
2. From the repro branch:

   npx -p hyperframes@0.8.4 -p @hyperframes/gcp-cloud-run@0.8.4 \
     hyperframes cloudrun render . --width 1920 --height 1080 --fps 24 --quality draft --wait

3. Read Plan.DurationMs from the workflow execution result (see branch README).
4. Repeat against a build that includes the single-pass scrub (or a 1-video control on the same unpatched image).

Expected behavior

Probe cost scales with duration, not V × D. On Cloud Run, plan time for N auto-start videos should stay near a 1-video control when seek work is the same.

Actual behavior

Unpatched coarse seeks for this fixture: 5 × 364 / 0.1 = 18,200 vs ~3,640 single-pass.

GCP Cloud Run results (same fixture: 31 scenes, 364s, 5 auto-start videos, nested GSAP with transform/blur weight, dummy color clips):

Image	Plan.DurationMs	Job wall
Unpatched main (~0.8.4)	10590	~163s
Build with single-pass scrub	5068	~70s

About 2× plan time on unpatched. This run did not hit protocolTimeout. Dummy 1s color clips still make each seek cheap; production hangs we hit used real 1080p footage and heavier per-seek GSAP, and failed the probe on Cloud Run.

Earlier empty-timeline fixtures (60s×6, 600s×6 with plain tl.set) showed almost no Plan gap — seek count alone is not enough on 8 vCPU. The tween weight in this branch is what made the A/B readable.

Environment

hyperframes doctor

  ✓ Version          0.8.4 (latest)
  ✓ Node.js          v24.19.0 (darwin arm64)
  ✓ CPU              0 cores · Unknown
  ✓ Memory           16.0 GB total · 2.6 GB available
  ✓ Disk             341.0 GB free
  ✓ Frames cache     /var/folders/bl/c4hk6r7n34g1d81txd99gsgr0000gn/T/hyperframes-extract-cache-501 · 341.0 GB free at /var/folders/bl/c4hk6r7n34g1d81txd99gsgr0000gn/T/hyperframes-extract-cache-501 · default
  ✓ Archive extractor unzip
  ✓ Environment      non-TTY
  ✗ whisper-cpp      Not found (optional — needed for transcription)
                     brew install whisper-cpp
  ✗ TTS (Kokoro)     Not installed (optional — local voice fallback)
                     pip install kokoro-onnx soundfile
  ✗ BGM (MusicGen)   Not installed (optional — local music fallback)
                     pip install transformers torch soundfile numpy
  ✓ FFmpeg           ffmpeg 8.1.2 at /opt/homebrew/bin/ffmpeg
  ✓ FFprobe          ffprobe 8.1.2 at /opt/homebrew/bin/ffprobe
  ✓ Chrome           cache: /Users/val/.cache/puppeteer/chrome-headless-shell/mac_arm-152.0.7977.42/chrome-headless-shell-mac-arm64/chrome-headless-shell
  ✗ Docker           Not found
                     https://docs.docker.com/get-docker/
  ✗ Docker running   Not running
                     Start Docker Desktop or run: sudo systemctl start docker

  ◇  Some checks failed — see hints above

Cloud Run A/B above used GCP Cloud Run + SwiftShader (hyperframes cloudrun), unpatched ~0.8.4 vs a build that includes the single-pass scrub.

Additional context

• Trigger: in-scene <video> with no data-start → data-hf-auto-start → visibility scan in browser_probe / plan.
• Do not treat a fast Mac render as a counterexample. The failure mode is Cloud Run plan/probe cost.
• A unit test that counts seekTl calls (V × samples vs samples) would pin the bug without relying on wall clock.