Myco Brain 1.2.6

[1.2.6] — 2026-06-17

An agent-experience release: the agent contract now teaches Myco as the
adjudication engine it is, and the doctor verifies your setup for real instead of
trusting config. No tool-contract changes (the brain_* inputs/outputs are
unchanged; brain_ingest gains an additive extraction field). No migrations.

Changed

• Agent contract rewritten to embody "the program writes the facts, not the
  LLM." The runtime contract, the pasteable CLAUDE.md / .cursorrules / AGENTS.md
  manual, the brain_* tool descriptions, and the website now teach the
  source-first write ladder (ingest a source -> propose a claim -> private
  save_memory) and the engine's mechanics: compounding confidence across
  independent sources, supersede-don't-overwrite on conflict, provenance via
  brain_why, and propose -> review -> promote. brain_save_memory is honestly
  fenced as a private scratchpad, never workspace truth. The runtime contract is
  single-sourced so the surfaces cannot drift (CI guard: npm run check:contract-drift).

Added

• mycobrain-doctor now verifies live, not just config. For the local Ollama
  path it pings the server, confirms the model is pulled, and runs a real embed
  and generation ("live-verified"); it adds an extraction-backlog check and a
  --fix mode that offers to pull missing models. brain_ingest returns an
  honest extraction: "graph" | "search-only" receipt so an agent knows whether
  a fact graph was actually built from the source.
• Agent-contract eval harness (npm run eval:contract, gated on
  ANTHROPIC_API_KEY): scores whether an agent routes to the right tool across 20
  adversarial scenarios, so the contract is measured and regression-guarded.