docs: sync README with research-assist tool, judge calibration, and 8-title status

[serpeigd]

What changed

Documentation-sync pass on the README (and two small doc/docstring fixes elsewhere). No code behavior changed; pytest tests/ still 8/8 green.

• Live demo link was missing entirely — added https://twistify.onrender.com with the already-documented Render free-tier cold-start/filesystem-reset caveat (couldn't independently verify reachability from this sandbox — outbound egress to that host is blocked by the sandbox's network proxy; treat as expected Render free-tier behavior per CLAUDE.md, not a broken link).
• Fixed a stale "7 hand-researched" count → 8. Gone Girl was added to content/researched/ (git log confirms, and the folder has 8 files) before this sync started, but the README, and docs/DESIGN.md's "Pending" section, and webapp/resolve_tmdb_ids.py's docstring, still said 7. Fixed all three. (CLAUDE.md itself has the same stale "7 hand-researched films" line in its "What this is" section — left untouched since it wasn't in scope, flagged for the repo owner instead.)
• Documented webapp/research_assist.py (D14) for the first time in the README: what it does (Wikipedia+TMDB retrieval → drafted ContentPack), the best-of-3 generation design, the sanitize_grounding() code-level safety net, and its honest current status (tested end-to-end on one title, Citizen Kane).
• New sections: Configuration (.env keys — TMDB/GROQ/ANTHROPIC/UPSTASH, all optional, nothing paid required), Project layout, Roadmap (mirrors CLAUDE.md's "Next task"), Limitations (states the SubstringJudge/LLMJudge recall gap plainly up top instead of leaving it implicit further down the page).
• Badges: added Groq and TMDB — both genuinely load-bearing (free baseline generator, LLMJudge calibration, research-assist, browse-tier posters) and previously undocumented.

Verified before writing

• git log --oneline -30, full repo tree, docs/DESIGN.md D1–D14.
• pytest tests/ -v → 8/8 pass.
• Ran python webapp/app.py locally, confirmed it serves and the two existing screenshots still match the UI mechanics shown (spoiler curtain, catalogue sidebar) — the sidebar's "Con contenido" badge count in the screenshots is stale at 7 vs. the real 8 now, a cosmetic one-off discrepancy not worth a screenshot regen per the task's "don't fabricate screenshots of features that don't visually differ" guidance; flagging it here instead.
• Every CLI command referenced in the README against the actual argparse/docstrings in evals/run_eval.py, webapp/prewarm_translations.py, webapp/resolve_tmdb_ids.py, evals/calibrate_*_external.py.
• .gitignore to confirm content/_drafts/ (research-assist output) and .env are correctly excluded, matching what the README now claims.

Left for the repo owner

• The LLMJudge/SubstringJudge recall gap (0.089 / 0.0) is a genuine, unresolved limitation — now stated plainly in a dedicated Limitations section rather than only inside the results tables. Per CLAUDE.md's "Next task", Milestone 1 (retrieval) shouldn't start until this is resolved.
• CLAUDE.md's own "What this is" section still says "7 hand-researched films" (its "Status" section elsewhere correctly says 8/20) — a small internal inconsistency, not touched since CLAUDE.md wasn't in this task's scope.
• Screenshot refresh (7→8 "Con contenido" badge) is optional/cosmetic, not blocking.

---

Generated by Claude Code