https://github.com/heygen-com/hyperframes/issues/1653 - still reproducible in 0.75.0

[yahelshiran-rvr]

Describe the bug

#1653 - still reproducible in 0.75.0

Link to reproduction

#1653 - still reproducible in 0.75.0

Steps to reproduce

#1653 - still reproducible in 0.75.0

Expected behavior

#1653 - still reproducible in 0.75.0

Actual behavior

Production data on @hyperframes/producer@0.7.5 per-frame regression persists

Reran the full bench-39 suite (38 templates) on 0.7.5. The metric split from #1657 is working ‚captureSetupMs is consistently low (1‚323ms). But captureAvgMs at p50 = 162ms is still ~2.5X the 0.6.42 baseline of 64ms. The regression is in captureFrameMs, not setup.

Summary stats

Stat	captureAvgMs
min	36
p5	74
p50	162
avg	203
p95	562
max	585

15 Examples

capture-regression-0.6.42-vs-0.7.5.csv

Environment

• @hyperframes/producer@0.7.5
• Node.js v22.16.0 (linux amd64)
• Chromium 149.0.7827.53 (bundled)
• FFmpeg 5.1.9-0+deb12u1
• AWS c6a.4xlarge (16 vCPU, 32 GB)
• fps=15, resolution=1920√ó1080 (except where noted), workers=4
• HF_STATIC_DEDUP_VERIFY ‚Äî unset (default)
• All times in ms

Environment

https://github.com/heygen-com/hyperframes/issues/1653 - still reproducible in 0.75.0

Additional context

No response

[miga-heygen]

Hi — I'm working on a fix for this regression in PR #1718. Two optimizations targeting the init-time overhead:

1. Synchronous GSAP proxy flush — exposes flushPendingOperations() as window.__hfFlushSync so headless renderers drain the tween queue instantly instead of waiting for rAF-based batch ticks (100 ops/tick at 33ms intervals). Also force-publishes the "timelines built" readiness signal immediately.

2. Parallel init polls — pollVideosReady, pollImagesReady + decodeAllImages, document.fonts.ready, and waitForOptionalTailwindReady now run concurrently via Promise.all.

I benchmarked against the compositions in hyperframes-launches, but those top out at ~220 GSAP tweens — not enough to surface the regression. The rAF drain completes within a single batch tick during the warmup loop, so before/after numbers are within noise.

Could you share the bench-39 template suite (or a subset of the tween-heavy templates) so I can validate the fix against the workload that surfaced the original regression? The fix should show the biggest improvement on compositions with thousands of GSAP tweens, where the rAF drain takes seconds instead of milliseconds.