test(mcp): baseline official 2026-07-28 conformance #1661

Verified change

The official MCP conformance project merged several material checks on September 6–7, 2026:

• PR #488: client authorization must send the RFC 8707 resource value exactly as published by Protected Resource Metadata; normalizing a bare origin by adding / can break exact-match authorization servers.
• PR #466: the runner now sends resource in both authorization and token requests when configured.
• PR #492: three consecutive tools/list calls on the 2026-07-28 wire must keep an unchanged tool set in deterministic order; divergence is a conformance warning that CI treats as failure.
• PR #468: the validator now accepts extension result envelopes such as Tasks.

Repository search found no pinned official conformance baseline in EventRelay.

Why this matters

Agent Factory/OpenProtocol currently have project-specific fixture tests, including the draft Skills harness in #1655, but no implementation-neutral proof of the underlying MCP wire contract. Auth-resource normalization and non-deterministic discovery can respectively break enterprise OAuth and tool/prompt caches even when local unit tests pass.

Acceptance criteria

• pin an immutable official conformance commit; do not consume floating main
• inventory the MCP client/server surfaces being certified and record exclusions
• exercise applicable 2026-07-28 server scenarios against fixture/local-only endpoints
• exercise client authorization with a bare-origin PRM resource and prove exact preservation in both authorization and token requests
• prove stable tools/list order across at least three unchanged probes
• accept versioned extension result envelopes without weakening core JSON-RPC validation
• retain explicit coverage for pre-2026 dated-wire notification behavior where supported
• store a machine-readable receipt: conformance commit, implementation commit, scenario/check IDs, versions, warnings, failures, and skipped reasons
• fail closed on required checks; do not hide warnings in aggregate success
• run without production endpoints, secrets, paid providers, merge, or deployment

Relationship to existing work

This complements #1655; it does not replace Skills-specific origin, digest, collision, or approval-revocation checks.

Estimated effort

1–2 engineering days for a pinned local harness and baseline receipt.