D128 — Serve the 2026-07-28 revision alongside the handshake era

(Superseded in part by D168: the decision to serve both eras stands, but it is the official SDK that implements it — the era resolution, envelope and header rules described below are no longer Cartographer code.)

Decision. The server answers both protocol generations at once and decides which one applies per request: a request is 2026-07-28-era iff params._meta.io.modelcontextprotocol/protocolVersion is present or the MCP-Protocol-Version header is (Request.resolveEra, internal/mcpserver/protocol.go), everything else is the handshake era. The era's result envelope (resultType: "complete" plus _meta.io.modelcontextprotocol/serverInfo) is applied in exactly one place, on the way out of dispatch, so no handler knows about eras; tools/list additionally carries ttlMs and cacheScope: "private" in the new era. server/discover is implemented and answers in both eras; initialize and ping stay routed for the handshake era. Over HTTP the three mirror headers are validated against the body (-32020, HTTP 400), an unsupported version is -32022 + 400, an unknown method is 404 in the new era and 200 in the handshake era, and GET/DELETE on the MCP endpoint are 405. Origin is validated on the MCP endpoint against an operator allow-list that defaults to same-origin, before authentication runs, and the CORS allow-origin header echoes the accepted origin instead of *. The resources capability, which no method ever backed, is no longer advertised. The CLI client's reachability probe moves from ping to server/discover and now surfaces the JSON-RPC error carried by a 400 or 404 instead of the bare status.

Rationale. What 2026-07-28 removes — the handshake, sessions, SSE resumability — is what this server never had, so the architecture already matched the new model and only the protocol surface lagged. Flipping the version outright was not defensible: the actual consumers are Claude Code, Codex, Kiro and OpenCode, and it is not established that any of them speaks the revision yet, so a flip would have broken every live KB client to satisfy a spec no client had asked for. Serving both is cheap precisely because the server is stateless — the era is a property of a request, not of a connection, so coexistence costs no bookkeeping. The Go SDK was rejected: adopting it would mean rewriting dispatch, the policy gate, the audit wrapping and the tool-prefix machinery, all built around the hand-rolled Request/Response, to close a conformance delta of a few hundred lines — and it would add this project's first protocol dependency. The mirror headers are validated rather than trusted because they are client-controlled and duplicate what the body already says; letting an authorization decision read them would create exactly the split source of truth the spec mandates the validation to close. Origin validation has been a MUST since 2025-03-26 and was simply missing; a wildcard allow-origin alongside it is what turns a rebinding attack into a readable response.

Consequences. Existing clients are unaffected: a handshake-era request produces byte-identical responses, which is asserted directly (TestHandshakeEra_ResponsesUnchanged) and indirectly by the e2e suite passing unedited. Operators who expose the server to browsers must now set mcp.allowed_origins, or accept the same-origin default — the one behaviour change that can turn a working setup into a 403, which is why "*" exists as an explicit opt-out. The same-origin default is not by itself a rebinding defence, since a rebound name matches both Origin and Host; the allow-list is. resources disappearing from the capability map costs nothing observable, since a client acting on it got -32601 anyway. skills.listChanged stays, because notifyWrap and artifactNotifyWrap do emit that notification — an earlier reading that called it unbacked was wrong. The CLI client deliberately stays handshake-era, _meta-free: the server serves both, and moving the client only pays off once the handshake era is retired, which is deferred to D130 and gated on the per-request era roster of D129.