spoolcast dev-log v2 shipped — rules that can say no, audit gates the render

second spoolcast video live. topic is the pattern i kept hitting in long-running AI sessions — i set a rule, the AI breaks it, then quietly rewrites the rule so the break looks intentional. after nine takes of a single video's reveal animation, the real problem wasn't the AI's output. the rules file had mutated every session and i'd been approving each edit without noticing. the fix is one sentence: the AI surfaces rule conflicts, shows three options (update / exception / decline), waits for the human.

the production itself caught a bunch of pipeline bugs i didn't know existed:

• white-flash gaps at every chunk boundary — traced to chunk_end_frame not updating after late-stage duration bumps (high-weight tail pauses, hold_duration_sec extensions). 28 gaps across the video, each a 0.2-0.7s blank white frame.
• paint-on stroke-reveal starting from a white canvas. deferred the whole paint-on entrance until the preprocessor outputs RGBA frames where unpainted pixels are transparent.
• crossfade underlays rendering at 1.0 zoom when the prior chunk was mid-zoom-in. snap-back mid-fade. fixed by applying the prior chunk's final camera state to the underlay.
• priorFinalImageSrc pointing at the same image for every chunk because a loop refactor dropped enumerate and i leaked from a prior loop scope. every crossfade showed the wrong underlay (the last chunk's image) for a session before i caught it.
• thumbnail letterboxing on youtube — kie.ai outputs at 1376x768 (1.79:1) but youtube expects exactly 16:9, and adds a thin black bar to pad. fixed by rescaling to 1920x1080 before upload.

by the end built audit_render.py as a gate. extracts frames at every chunk boundary from the final mp4 and flags any that compress below a 30KB threshold (pure white 1920x1080 jpg is ~13KB). rule: render doesn't count as shipped until the audit passes. works in human-in-loop and autonomous modes — the audit is the verification, not code changes or eyeballing a few frames.

also landed this round:
- overlay-form meme punchlines — REJECTED stamp drops onto the prior illustrated scene at a rotated angle instead of replacing it full-frame. same for the no-AI icon. the stamping ON the scene reads harder than substituting the scene.
- readtime_override flag for author by-ear hold-duration tuning that bypasses the read-time validator with intent recorded, so editorial decisions aren't silent overrides.
- SRT captions now include bracketed [on-screen: ...] cues for text rendered on the frame — sound-off viewers see every rendered word, not just narration.
- copy-principle guideline: every title should hit something in the human psyche, not describe the topic. "sell the cheat, not the lesson" is one pathway; naming a specific private frustration is another. the meta rule is "hit the drive, not the subject."

the meta-lesson from making the video: "verified" has to mean the audit passed on the final mp4. not "code changed," not "preview-data looks right," not "i extracted a few frames." at one point i'd marked five fixes as done based on the intermediate data, rendered, and the user watched five seconds before seeing every flash still there. the mechanical check is what converts "i probably fixed it" into proof.

youtube: https://youtu.be/i3Z480n1k6k