MCP 2026-07-28 (stateless) spec support

[quinnj]

Summary

Implements the newly released MCP 2026-07-28 spec (announced today) as a dual-era implementation: modern clients are served statelessly per the new revision, while existing 2025-11-25 clients keep working unchanged on the same endpoint.

Stacked on #4 — review after that merges (this PR's base is spec-2025-11-25-fixes; retarget to main once #4 lands).

Server

• Stateless request handling: requests declaring 2026-07-28 (via params._meta io.modelcontextprotocol/protocolVersion or the MCP-Protocol-Version header) bypass sessions and the initialize handshake entirely. Era selection follows the spec's dual-era server rules.
• server/discover with supportedVersions, capabilities, instructions, and _meta serverInfo (cacheable).
• Header routing: validates Mcp-Method on every request and Mcp-Name on tools/call/prompts/get/resources/read against the body, including the =?base64?...?= sentinel encoding → HeaderMismatch (-32020); unsupported versions → UnsupportedProtocolVersionError (-32022) with the supported list; removed methods (ping, initialize, logging/setLevel, resources/subscribe, ...) → 404/-32601 on the modern path; resource-not-found aligned to -32602.
• resultType on all modern results; ttlMs/cacheScope on the five cacheable results (configurable via cache_ttl_ms/cache_scope).
• MRTR: handlers return MCPInputRequired(input_requests, request_state); input_responses(context)/request_state(context) expose the client's retry data. This is the protocol-native fit for flows like AiAPI's confirm-booking explicit-confirmation step.
• Request-scoped notifications: send_progress!/send_log! (gated on progressToken / io.modelcontextprotocol/logLevel _meta per spec) are delivered as an SSE response stream ahead of the final response. Handlers deliberately run on the request task so task-local auth contexts (e.g. AiAPI's Roam context) survive.
• subscriptions/listen: long-lived per-request SSE streams with the acknowledgment notification, subscriptionId tagging on every delivered notification, opt-in filters (toolsListChanged, promptsListChanged, resourcesListChanged, resourceSubscriptions), and graceful closure (close_subscription_listeners!). Wired into the existing list-changed/resource-updated broadcast helpers.

Client

• protocol_version >= "2026-07-28" switches the client to modern mode: no handshake, _meta (protocolVersion/clientInfo/clientCapabilities) injected into every request, Mcp-Method/Mcp-Name headers sent.
• initialize_client! transparently becomes server/discover for modern clients (drop-in for existing call sites); discover_server_info! exposed directly.
• MRTR: is_input_required(result) + input_responses/request_state kwargs on call_tool/get_prompt/read_resource; meta kwarg on call_tool for progressToken/log-level opt-in.
• listen_subscriptions! opens the long-lived stream and dispatches notifications to register_notification_handler! handlers in real time.

Not in scope (follow-ups)

• x-mcp-header custom tool-parameter header mirroring (client MUST per spec — needs tool-schema awareness at call time)
• True incremental streaming of in-flight notifications (currently buffered until the handler returns, wire-format correct)
• CIMD client registration (DCR deprecation) — lives in OAuth.jl

Test plan

• Full suite green on Julia 1.12.6 — 227 tests, including the new 58-test MCP 2026-07-28 set: stateless request/validation matrix, MRTR round trip, modern-client e2e over live HTTP (discover → list → call → streamed progress/log notifications → subscriptions/listen ack/notify/graceful close)
• All legacy (2025-11-25) testsets unchanged and passing — dual-era behavior verified
• JuliaC --trim=safe harness still green (0 verifier errors/warnings)

🤖 Generated with Claude Code

[quinnj]

CI is fixed — two distinct issues, both only reproducible under HTTP.jl 2.x (CI resolves 2.5.5; the local dev manifest had 1.11.0 which masked them):

1. World-age: handlers registered after HTTP.serve! starts (like the test's noisy tool) failed with "method too new to be called from this world context" when connection tasks dispatched them. All user-supplied handlers (tools/prompts/resources/completion/logging/cancellation/request hook) are now invoked via Base.invokelatest — standard practice for servers supporting runtime handler registration. This was a latent bug on main too; no prior test ever called a late-registered tool.

2. HTTP.jl 2.x server buffers SSE bodies until stream close — verified against HTTP.jl's own canonical sse_stream do-block example (curl -N receives all events in one burst; HTTP 1.11 streams them live). Protocol-wise subscriptions/listen still delivers everything correctly in order at graceful closure, but real-time delivery needs a fix in HTTP.jl itself (filed separately). The subscriptions test is now transport-aware: it asserts live mid-stream delivery under HTTP 1.x and complete delivery-at-close under both.

Full suite verified locally under both HTTP 1.11.0 and HTTP 2.5.5 (fresh-resolve env mirroring CI).