Illustration register: toy diorama -> forensic-reconstruction CGI

[upgradedev]

Summary

The generated illustration still and clip currently read as a toy/game
(literally "Miniature diecast toy car diorama on a printed road play mat"),
which undercuts the product's credibility with judges and insurers. This
raises production quality while keeping the output unmistakably a synthetic
reconstruction, never crossing into fabricated-evidence territory.

• src/claimscene/report.py: _diorama_scene_description renamed to
  _forensic_scene_description; illustration_still_prompt() and
  illustration_prompt() rewritten into a clean, computer-generated 3D
  forensic accident-reconstruction register (accurate vehicle proportions,
  real road surface and lane markings, neutral daylight, professional CGI
  clarity). Still built only from the closed vocabulary, still
  deterministic, clip prompt still ends with Overlay text: '{DISCLOSURE}'.
• The prompts state plainly the output is not a real recording, and
  never use a word that could nudge the generative model toward
  photorealism (photorealistic, photograph, dashcam, real footage,
  cinematic film still, documentary footage are all absent, verified by
  a new test in each of tests/unit/test_report.py and
  tests/integration/test_pipeline.py).
• scripts/readiness.py's check_genblaze_illustration_port_sealed now
  checks the not-a-real-recording marker in both the still and clip
  prompt fields (previously only checked the clip prompt) - strictly
  stronger than before.
• README.md (the "illustration layer" bullet) updated to describe the new
  register.

Untouched (by design)

• DISCLOSURE / WATERMARK constants, the clip's overlay-text instruction,
  provenance.py, everything under tests/security/.
• The eval-set descriptions (README ~312/332, generate_eval_scenarios.py,
  eval/scenarios/manifest.json) - those describe the staged synthetic
  input photos used for the extraction-accuracy eval, a separate concern
  from the illustration output register, and remain accurate as written.
• eval/evidence/live_illustration/ (a hash-pinned real-provider run
  captured under the old register), frontend/src/assets/hero-illustration.jpg
  • its alt text in Hero.tsx, and frontend/src/test/fixtures/golden-manifest*.json
  • all predate this change and are either hash-pinned or would need a real
    paid Genblaze call to regenerate; none are in scope here. The committed
    demo video (demo/claimscene-demo.mp4) also predates this change and
    still shows the old toy-diorama illustration.

Test plan

• pytest tests/unit/test_report.py tests/integration/test_pipeline.py -v - 17 passed
• pytest --cov=src --cov-fail-under=90 tests/ (scoped around a pre-existing,
  environment-only pytest collection quirk on this machine, confirmed
  identical on bare main and confirmed green in the last real CI run
  on this exact commit) - 236 passed, 94.64% coverage
• python scripts/readiness.py --min 95 - 100% automatable, gate PASS
• ruff check src tests scripts - clean
• CI (this PR) - all 9 jobs green