v0.3.6 — Pipeline quality fixes + Opus 4.7

Changes in v0.3.6

Pipeline quality improvements

• Abstract-only source cap (30%): Drops lowest-scored abstract-only papers when full-text alternatives are available. Addresses weak evidentiary base complaint.
• Prestige-term downgrade: "systematic mapping" → "narrative mapping", "composite scoring" → "weighted ranking", etc. Prevents overclaiming rigor in narrative reviews.
• AI-assistance disclosure: New Methodology section 2.10 explicitly documents LLM role and limitations.
• Extended numeric consistency validator: 4 new regex patterns catch corpus count contradictions ("a total of N", "corpus comprised N", etc.).

Model support

• Claude Opus 4.7 added to Anthropic model dropdown (latest flagship).
• Gemini 3 Flash Preview available in Google model list.
• gemma4:26b added to Ollama model list for remote server use.

Bug fixes

• Gemini NoneType crash when content.parts is None (streaming + non-streaming paths).
• Counter-evidence search query builder no longer produces malformed queries like "efficacy of limitations".
• claim_ledger max_tokens increased 16K→32K to prevent JSON truncation on thinking models.
• LLMResponse.strip() bug in citation fixes — now extracts .text attribute correctly.
• Section generation retry on empty responses.
• All LLM backend timeouts raised 600s→900s (Google from 300s).

Writing prompts

• Bracket citation format [Author, Year] now consistent across all section templates (previous versions had contradictory examples).
• Word targets raised: Intro 700→900, Related Work/Results/Discussion 1400→1800, Conclusion 350→450.
• Removed paragraph-count ceiling (was "9 paragraphs of ~150 words each") that caused sections to stop at ~1000 words.
• Citation-breadth rule: prompt now encourages citing ≥70% of provided references.