feat(illustration): seed the clip from the deterministic schematic

[upgradedev]

Summary

• The illustration clip was seeded by a text-to-image generated still. A live render on 2026-07-30 (case live-forensic-retry-0bb32eeb) proved prompt-level control alone cannot pin a diffusion model's geometry: even with an explicit "the red car is directly behind the blue car, nose to tail" sentence in the prompt, the model rendered a head-on collision for a rear-end fixture.
• The clip is now seeded by a raster of the case's own deterministic schematic (SchematicArtifacts.seed_png, the impact-frame render with all drawn text/watermark omitted via render_frame's new annotate flag), so the clip's geometry is inherited from the factual layer instead of requested from a generative model.
• The clip prompt (report.illustration_prompt) is rewritten into a top-down-diagram register ("animate this top-down forensic reconstruction diagram... vehicles stay perfectly static... only the camera moves"), matching what the seed image actually looks like, while keeping the honesty markers (computer-generated, not a real recording), the no-on-image-text instruction, and the geometry sentence.
• The manifest's illustration.still_model / illustration.still_prompt no longer claim a text-to-image model produced the seed (still_model is honestly null); a new illustration.still_source: "schematic:impact_frame" field names the seed's real provenance. Existing structural checks (provenance.verify_all, scripts/readiness.py) keep working unmodified against the corrected values.

Test plan

• pytest --cov=src --cov-fail-under=90 tests/ — 454 passed, 1 skipped (live-VLM, no creds), coverage 96.95% (up from 94.84% baseline)
• ruff check src tests scripts — clean
• python scripts/readiness.py --min 95 — PASS, automatable 100.0%
• New test asserts the clip provider's input bytes equal the independently-recomputed schematic seed (not just "some input")
• New test asserts burning the disclosure onto the seed reproduces the sealed illustration_still bytes exactly (single-caption invariant)
• Updated tests/integration/test_illustration_watermark_burn.py to reflect that the still burn no longer depends on the (possibly misbehaving) clip provider
• CI (all jobs) — polling after this PR opens