docs: sync README with closed judge search, Milestone 1, and a legal notice

[serpeigd]

What changed and why

Scheduled documentation-sync run. README.md only — no product or measurement-track code touched.

The README had drifted behind D15/D16. Its Roadmap still listed "Resolve the judge-trust gap before Milestone 1 (retrieval)" and its Limitations still said "this is why Milestone 1 hasn't started" — but Milestone 1 is complete at 20/20 titles, and the judge search has been closed after six attempts.

Stale numbers, verified against the tree

Claim	Was	Now
Test suite	8 tests (4 places)	26 — counted across the five files in tests/
Design log	D1–D14 (2 places)	D1–D16
judge.py	"SubstringJudge + LLMJudge"	all six judge classes

Status / Roadmap / Limitations rewritten to match reality

• Judge calibration is now marked closed-and-unsolved rather than in progress, with the reason SubstringJudge stays the default stated plainly: not because it's good, but because missing things is a bounded failure mode and inventing them isn't. The SimilarityJudge result that ended the search is included — a film's own public premise scored 0.561, higher than the one confirmed real leak at 0.525.
• The three confirmed leaks are documented as a limitation, with both mechanisms kept distinct: Los cronocrímenes leaked from the model's own memory (the corpus never mentions it), while Tetsuo and Hard to Be a God leaked because Wikipedia's own overview section carries those films' core spoilers verbatim — the GREEN corpus itself wasn't safe for them. Come and See has the same risk latent. Milestone 1's real gain is stated alongside it (grounding 0.0 → 1.0, judge-independent; the five highest-risk twist titles clean) without letting it read as a safety guarantee.
• Roadmap replaces the obsolete judge-then-retrieval sequencing with what actually replaced it: --save-briefs plus a human reading the text, which caught all three leaks when no judge caught any.

Project layout

Filled in the modules added since the last sync: retrieval.py, retrieval_groq.py, retrieval_prompts.py, stats.py, external_dataset.py, train_spoiler_classifier.py, generator.py/demo_generator.py, and the three newer calibration scripts.

Legal notice (requested directly)

## License previously read "No license defined yet — personal portfolio repo." It's now an explicit notice: copyright reserved, with the point that public visibility on GitHub is not itself a licence, plus the third-party terms this repo can't relicense — TMDB's non-commercial attribution terms, Wikipedia's CC BY-SA, the IMDb Spoiler Dataset (gitignored, never redistributed; direct scraping prohibited by ToS and not done), and the cited director/critic quotes in content/researched/*.json.

No LICENSE file was added — choosing a licence is the author's call, not something a docs-sync run should decide. Flagged below.

Verification

Counts and module lists come from the tree itself (tests/, evals/judge.py, docs/DESIGN.md, run_eval.py's argument parser), not from CLAUDE.md. pytest isn't installed in this environment, so the suite wasn't run locally — CI is the check.

Needs a human decision

Whether to add an actual LICENSE file (and which one). The README now states the default reservation accurately in the meantime.

---

Generated by Claude Code