[yia-mw-agent]

Tracking issue: MachineWisdomAI/fava-trails.org#36 (product companion; the dependent website PR is opened after this one and merges after it).

What changed

Integrates the accepted agent-entry package (artifact accepted-agent-entry-package.md, SHA256 7e910c5f4b8dbff71d2f91a951bfae4eb0aff5c905668d27827497e444882f0a, editor revision 1431) exactly as accepted, with the binding coordinator decisions applied.

• AGENTS_USAGE_INSTRUCTIONS.md: full replacement. Adds the rejected/errored revision path (fresh save_thought with parent_id; supersede again for a failed replacement successor), the distinction between a persisted verdict (thought.validation_status is rejected or error) and pre-review failures that leave the draft unchanged, one governed-recall statement with a link to docs/governed-recall.md, conditional pointers to setup/README/CONTRIBUTING, and removes the legacy memory/shared and memory/branches fallback.
• AGENTS.md: tool-reference patches A1-A10 (contributor workflow moved out, parameter rows and propose_truth/supersede/sync text carry the same branches, error listed as a frozen status).
• CLAUDE.md: short adapter pointing at the owners.
• CONTRIBUTING.md: new "Working in this repository" owner section, machine-specific path removed.
• AGENTS_SETUP_INSTRUCTIONS.md: S1-S10 (install order pip, version, install-jj; release-neutral version framing; absolute FAVA_TRAILS_DATA_REPO before doctor; credential source instead of a literal key; fails closed only on the effective configuration; recorded-run evidence statement; registration without a pointer loop; destination plus routing qualification; stable-role identity on a second machine).
• README.md: R1-R2 remove the claim that setup covers Claude sections. Windows/WSL, Codex and cleanup procedures are unchanged (the diff touches two sentences only).
• docs/runtime-and-upgrade.md: T1-T3 release-neutral intro, 0.7.0 kept as a dated example, install command no longer pins >=0.7.0.
• docs/mcp-context-overhead.md: measured get_usage_guide size updated from 2965 tokens / 11860 chars to 3737 tokens / 14947 chars (measured with fava-trails measure-mcp-context --surface full). The test was not changed or weakened.
• Version: scripts/mw-version.py update --base origin/main bumped 0.9.1rc1 to 0.9.2rc1, because AGENTS_USAGE_INSTRUCTIONS.md is a release_path. mw-version.py check passes. CHANGELOG Unreleased entry added.

No package, tag or release was created; this is source only. The 0.9.0rc1 candidate walkthrough is referenced as one observation, not a benchmark.

Verification

• pytest: full suite 1233 tests pass (1230 in the main run; the three tests/test_packaged_mcp.py tests need the uv toolchain and passed on rerun with uv installed). Ruff reports two findings, both in scripts/mw-version.py (not changed here), and they also appear on origin/main.
• Relative link and anchor check over AGENTS.md, USAGE, SETUP, CLAUDE.md, CONTRIBUTING, README, runtime-and-upgrade: 108 links, 1 unresolved: ../src/fava_trails/protocols/secom/README.md in SETUP, which is pre-existing on origin/main and out of scope.
• Credential grep (=placeholder, $(cat, --api-key, export ...KEY=) over SETUP, README, USAGE: no matches.

Scenario evidence (evidence classes kept apart)

• A, reachability (script over the integrated text, no model): 38 of 39 assertions pass. The 39th is a regex false positive for "rank" matching "reranking" in an existing Machine Wisdom link description in the website llms.txt; the file makes no ranking claim. Covers: install order and doctor with an absolute data path (S1); Target thought is unavailable to this caller, fresh parent_id draft, persisted-error needs a returned thought.validation_status: error, pre-review failure leaves the draft, Jev probability only (S2); default recall approved-only, mode="authoring" identity-scoped, history needs an operator endpoint, lexical recall, list_scopes before guessing, advisory is not human approval or fact verification (S3); supersede successor, original stays current until approval, supersede again after a rejected successor, never delete .jj/fava-governance-transaction.json, successor carries its own evidence (S4).
• B, disposable repo with an isolated XDG_CONFIG_HOME and a placeholder endpoint that was never contacted (S1), run against this branch's source build 0.9.2rc1 (editable install; not a release): bootstrap with no reviewer options wrote only trust_gate: llm-oneshot and doctor stopped with guidance, nothing sent; an endpoint with credentials was refused (exit 2); bootstrap with all three options reported policy, provider, model, endpoint and loopback endpoint on this machine; doctor printed the effective policy, provider, model, destination local_endpoint, the credential variable name only, and API key: NOT SET; register --agent-id printed a registration with an operator-set FAVA_TRAILS_AGENT_ID and made no change.
• C, recorded real-model run: reused as is from docs/local-review-walkthrough.md (installed 0.9.0rc1 candidate, one local model, two rejections with reasoning, an approval and a correction). No new model run was made. Fixtures and the instruction checks above do not establish model quality.

Not done here (carry-forward for #38)

• Final supported version and publication wording, release notes, and the decision about the older docs/fava_trails_faq.md "Published 0.7.0" lines (archived bodies untouched).
• docs/runtime-and-upgrade.md still describes an upgrade baseline from published 0.6.0 further down (lines untouched here; check against .github/workflows/release.yml).
• CONTRIBUTING "Release to PyPI" steps 5-6 (vendor copy, ~/.claude.json) left verbatim.
• No live MCP/client configuration, identity, credential or stored-record change; no remote publication exercised.

[timeleft--]

Independent Captain review of exact head 26b1c4f: all 12 changed files reviewed against the accepted canonical package and owning source. Accepted New/full-file payloads match verbatim; README unique client and cleanup procedures remain intact. Actual approval/error schema, failed-successor retry lineage, stable-role identity, effective routing and credential-source qualifications retained. Version policy independently passes 0.9.2rc1; 24 focused MCP context tests pass, including shipped usage-guide measurement coupling. All 12 current-head hosted checks pass, including both required test runs and Validate PR app15368; no commit statuses or unresolved threads, fresh base f772243. Release/publication/live acceptance remain separate; existing bounded final-assembly carry-forwards remain with #38.