[recipes] Wiki pipeline: OpenRouter + thought_edges + prompt safety

[AnkitClassicVision]

Summary

Three related wiki-pipeline improvements, packaged together because they land as a coherent upgrade after the typed-reasoning-edges feature (#208) merged.

1. typed-edge-classifier: OpenAI-compatible LLM path

Previously required ANTHROPIC_API_KEY direct to api.anthropic.com. Now also accepts LLM_BASE_URL + LLM_API_KEY (OpenRouter, LiteLLM, etc.) — the same pattern the sibling entity-wiki and wiki-synthesis recipes use. Direct Anthropic remains the fallback when the OpenAI-compat vars aren't set, so existing installs keep working.

Why: running all three recipes through one provider avoids maintaining a second API key just for the classifier.

2. entity-wiki: thought_edges integration

Entity wiki pages now include a Thought reasoning links subsection under Relationships, rendering supports / contradicts / evolved_into / supersedes / depends_on / related_to edges between the evidence thoughts the page already cites.

Graceful degradation: if schemas/typed-reasoning-edges/ isn't installed (table missing), the fetch returns [] and the subsection is omitted — no runtime error.

Why: closes the loop opened by #208 — the classifier produces thought_edges rows; now entity-wiki actually renders them.

3. wiki-synthesis: bounded-size prompt + schema tolerance

• Autobiography builder caps each entry at 240 chars and each per-year sample at 25k chars (with a 25-entry floor so sparse years still synthesize). Prevents context blow-ups on long entries (meeting notes, code blocks in Obsidian captures).
• fetchThoughts gracefully handles deployments whose thoughts table lacks a source_type column by falling back to metadata inference (metadata.source_type / source_system / source / kind). Preserves server-side filtering when the column exists.

Provenance note

These changes were in my working tree when picking up wiki-compiler install work this week. I believe they came from an earlier Claude/Codex session that didn't land a commit — the code style and comment density match the rest of the repo, and the changes are tested end-to-end against a real Open Brain (2000+ thoughts, 5000+ entities). Opening as a PR so the maintainer can decide whether to accept or discard.

Test plan

• Classifier: run with OpenRouter credentials (LLM_BASE_URL=https://openrouter.ai/api/v1 LLM_MODEL=anthropic/claude-haiku-4-5) — 150 pairs processed, 4 edges inserted, $1.86 spent under $2 cap.
• Entity-wiki: generated 690 pages against a brain with 8 thought_edges. Pages with linked thoughts that have reasoning edges between them render the new subsection; others omit it cleanly.
• Wiki-synthesis: autobiography regenerated from 357 thoughts. Prompt size stayed under 25KB.
• Verified back-compat: classifier still works without LLM_BASE_URL set (falls back to Anthropic direct).

🤖 Generated with Claude Code