Vary SIM prose, domain-aware Python hypotheses, per-idea citations + simulated thinking

[duckyquang]

Problem

The keyless SIM generators (frontend/src/lib/sim/content.ts, webapp/content.py) were fixed-skeleton slot-fill: the self-critique / stress / review stitching sentences were identical every round and session, the entire research-proposal overview prose was byte-identical every session, and the Python make_hypothesis was always biomedical regardless of goal. Users saw the same text template over and over.

What changed

De-template (both runtimes, seeded r.choice([...]) banks; headings + structure unchanged):

• makeReview / make_review — the four dimension paragraphs + assumption note vary (scores drawn first, so they stay deterministic).
• makeSelfCritique / make_self_critique — fixed reread/doubt/stakes stitching now varies per round + target.
• makeStressReport / make_stress_report — break/claim/attack/feasibility/citation sentences vary per hyp + round (verdict token still seeded separately so report ↔ ranking agree).
• makeOverview / make_overview — every top-level section and per-proposal block drawn from banks, seeded per session (via proposal ids) so two runs of the same goal read differently; per-proposal streams keep blocks distinct. Prose genericised to fit any domain.
• Match rationales varied in engine.ts, simulator.py, seed.py.

Domain-aware Python hypotheses: ported the TS DOMAINS table + keyword extraction + title scaffolds into webapp/content.py, so a non-biomedical goal yields on-topic hypotheses. make_plan reports the inferred domain. Both DOMAINS tables enriched (reasoning/LLM terms → computing; inflammation/senescence → biomedicine).

Per-idea citations: fixed a browser bug where overviewRefs re-sampled random curated papers instead of each proposal's real citations (so the overview cited different papers than the drawer/donut). It now consumes p.citations, deduped like the Python _overview_refs, with [n] markers spread across the claim / why / experiment sentences. Overview ↔ drawer ↔ per-proposal donut now cite the same papers. Framing / summary / landscape / comparative / recommended / open-questions stay uncited (Co-Scientist synthesis). The Groq-prose path stays honestly uncited.

Simulated thinking: every sim hypothesis gets varied seeded reasoning (distinct per hyp), threaded through engine.ts (PlanHyp/hypContent/addHyp/toHypothesis) and the Python store (hyp_thinking side table, read in get_hypothesis). Groq GenHyp gains a real reasoning field (prompt + parse) that becomes thinking for BYOK hyps. types.ts Hypothesis + PlanHyp gain optional thinking.

Verification

• npm run build — green.
• pytest co_scientist/tests/unit -q — 276 passed.
• ruff — no new errors (net −3 vs baseline; the one unavoidable en-dash gets a scoped # noqa).
• Server E2E (two same-goal runs): prose differs between runs while ## headings match; per-proposal [n] resolve to the References list; framing/summary uncited; hypotheses on-topic (not biomedical); each hypothesis carries distinct thinking; self-critique rounds read distinctly.

🤖 Generated with Claude Code