Probe capability-gated session/close in the protocol matrix

[sweetu12]

Summary

• replace the stale session/stop capability, probe, JSON, Markdown, and summary fields with session/close
• gate close on a present object-valued sessionCapabilities.close advertisement and a successfully created real session ID
• execute close once and last, preserving unsuccessful JSON-RPC results instead of treating parameter or missing-resource errors as support
• emit probeSchemaVersion: 2 and reject reuse of every unversioned or non-v2 snapshot
• add deterministic standard-library tests and run them before the nightly production matrix step

Behavior

Negotiated close	Successful session/new with ID	Matrix action/result
omitted or null	either	no close request; not_applicable
advertised object	no	no close request; not_probed
advertised object	yes	one session/close, last, with the returned ID; preserve actual result

A successful close must contain an object result and no simultaneous JSON-RPC error. {} and an object with valid reserved _meta are accepted. invalid_params, resource_not_found, malformed or ambiguous responses, and other errors remain visible and do not increment the close-supported summary count. Ambiguous session/new responses likewise cannot supply a trusted session ID or schedule close.

Schema 2 is deliberately not backward-reused. With --changed-only, an existing unversioned/schema-1 or mismatched snapshot is discarded as a reuse source and every selected agent is probed again. Historical committed snapshots are not rewritten by this patch.

Test plan

• python3 -m unittest tests.test_protocol_matrix_close -v
• uv run --with pytest pytest tests/ -v
• uv run --with ruff ruff check .
• uv run --with ruff ruff format --check .

Local Windows evidence:

• focused standard-library suite: 18 tests passed
• Windows-compatible repository suite: 134 passed, 2 skipped, 4 Unix-only tests deselected
• Ruff lint and format checks passed
• Python syntax compilation passed
• an in-memory mutant that probed close despite omitted/null capability was killed by both negative subcases
• an in-memory mutant that accepted a response containing both result and error was killed by the malformed-response regression test
• an in-memory mutant that trusted an ambiguous session/new result and scheduled close was killed by the production-path regression test

The complete unfiltered suite was also run on Windows. Its only four failures were existing Unix assumptions: two POSIX executable-bit assertions and two select()-on-pipe tests. No live matrix, credentials, external agents, or model execution were used. Linux process-I/O behavior should be confirmed by normal upstream CI; this change leaves those Unix mechanics unchanged.

Migration and scope

The first scheduled run after merge emits a new schema-2 snapshot and fully re-probes selected agents. There is no stop-to-close row translation. This PR does not claim ACP conformance, modify agent metadata, rewrite historical generated snapshots, or add authentication/model-dependent lifecycle flows.