[abrichr]

Non-blocking notes from the delta review of #1542. No behavior changes.

Changes

• PROTOCOL.md: the renumbering note was inside the body of pre-registered Amendment 11. It is now its own short note before the amendments. It says old Amendment N is now N+1, and it names the three files that still use the old numbers: the two 2026-10-25-rb* run meta.json files and HAND_REVIEW.md. Those files are not edited. No amendment's rules changed.
• BrowserTeachBridge.hold_needed and VerifyBridge.hold_needed: the docstrings said the tool could be teach_run. Neither bridge offers it. They now say the bridge has no person holds and tool is ignored.
• BrowserTeachBridge class docstring: "none can change data" was stale since write contracts (feat: verified writes for any agent, and taught browser writes with receipts #1589). It now says that without a contract nothing can change data, and with one the wire check lets only the contract's writes out.
• MCP server instructions: the fallback for a teach bridge without its own instructions now names teach_run (web apps) next to teach_act.
• New tests/test_mcp_teach_bridge_contract.py: builds each real teach bridge (TeachBridge, BrowserTeachBridge, VerifyBridge) into build_server, calls teach_act and teach_run through an in-memory MCP client, and checks that hold_needed gets tool= with no internal failure. It also binds the signature for both tool names. With tool removed from BrowserTeachBridge.hold_needed, the test fails.

Finding: volatile fields in the read-back

New tests/test_readback_volatile_fields.py runs RecordReadBack on the real load_observer path against a local FHIR server whose resources get a new meta.lastUpdated and meta.versionId on every read.

• With the default field_paths, meta fields are not mapped, so the read-back never sees them. The save reads back as landed.
• If an operator maps meta.lastUpdated or meta.versionId in field_paths, run 1 stops as wrong_record, although only a read changed the starting record. This fails closed, but it is a false stop. records_drop_fields removes the field, and the save then reads back as landed.
• db_snapshot is not a read-back source. load_observer has no kind for it; only the contract frame uses it, and it has its own ignore_columns. So the read-back can't give a false stop there.

Follow-up: the read-back can give a false wrong_record on FHIR when a mapped field changes on every read. A possible fix is to refuse, or warn about, meta.* paths in a run-1 verifier unless they are in records_drop_fields. Nothing in this PR changes that behavior.

Tests

pytest -n 2 on the two new files plus test_verify_tools.py, test_browser_teach.py and test_mcp_teach_cli.py: 44 passed. Ruff check and format pass on the changed files.

[abrichr]

Coordinator review: diff checked. Engine changes are docstrings and MCP fallback text only (browser_teach.py, mcp.py, verify_tools.py); the rest is the PROTOCOL renumbering note and two new tests (bridge-contract build test, read-back volatile-field test). No behaviour change. CI green. Approve and merge.