P1 ViCLIP semantic pre-validator before heavy VLM escalation

[zedarvates]

Add a lightweight semantic pre-validation stage for generated video clips before invoking a heavier VLM.

Goal

Use a small video-text model (ViCLIP-L/14 or equivalent compatible implementation) to score semantic alignment between shot_spec text and the generated clip.

Target flow:
shot_spec -> generated clip -> lightweight semantic score -> PASS | ESCALATE -> heavy VLM/temporal validator

Requirements

• provider interface; no hard dependency on one checkpoint;
• cache embeddings/results by content hash;
• configurable confidence thresholds;
• retain evidence/run_id/provenance;
• never auto-approve a clip solely because semantic similarity is high if structural/temporal validators fail;
• escalation to LFM/other heavy VLM on low confidence, mismatch, multi-subject ambiguity or temporal defect signals.

Benchmark

Use a fixed set of existing StoryCore clips including:

• clearly correct shot;
• wrong/missing object;
• identity drift;
• camera mismatch;
• temporally unstable clip.

Measure:

• pre-validator latency;
• GPU/RAM/VRAM use;
• false PASS rate;
• false ESCALATE rate;
• percentage of heavy-VLM calls avoided;
• end-to-end render validation time.

Exit criteria

Demonstrate that the lightweight stage safely removes a meaningful fraction of unnecessary heavy-VLM calls while preserving validation quality on the benchmark.