fix(render): preserve and validate muxed video frames

[miguel-heygen]

Summary

• preserve encoded packet timestamps by removing output -r from every stream-copy mux/remux path
• validate video packet-count and true video-stream-duration parity after both local and distributed assembly
• bound full-stream ffprobe scans and cancel them with the render abort signal

Root cause

A 17.5-minute render captured all 31,527 frames, but the final MP4 retained only 162 video packets while full-length audio remained. The stream-copy mux paths still passed output -r; FFmpeg may drop or retimestamp packets to satisfy that requested rate even with -c:v copy. A zero exit status therefore did not guarantee a valid video stream.

Regression coverage

• reproduces the reported 31,527 → 162 packet mismatch in parity validation
• rejects true video-stream-duration drift without substituting longer container audio duration
• verifies abort kills a full-stream packet probe
• verifies local assembly rejects post-mux parity loss
• verifies a 30000/1001 single chunk assembled under a 30/1 plan preserves all packets and source timing
• pins the absence of output -r in audio mux and faststart stream-copy paths

Verification

• engine ffprobe/chunk encoder: 103/103 passed
• local assemble stage: 1/1 passed
• distributed single-chunk regression: 1/1 passed
• engine and producer typechecks passed
• producer test classification passed (57 files)
• lint, format, tracked-artifact, fallow, and typecheck pre-commit gates passed
• independent adversarial review: no remaining findings

The host FFmpeg is 4.2 and cannot run the pre-existing distributed CFR test because it lacks -fps_mode; the new focused distributed regression passes.