Public API for multi-node session rehydration (e.g. transport.adoptSession(id))

[technicallychudi]

Problem

The onsessionclosed docstring in webStandardStreamableHttp.d.ts (lines 57-67) explicitly describes a multi-node deployment pattern:

if you are handling HTTP requests from multiple nodes you might want to close each WebStandardStreamableHTTPServerTransport after a request is completed while still keeping the session open/running.

This is exactly the pattern we need — we run an MCP server on Azure App Service behind a Cloudflare Worker proxy, and every container restart (deploy) wipes in-process session state, forcing clients to re-handshake.

To rehydrate a session on a fresh transport instance, downstream servers currently have to reach through private fields:

const inner = (transport as any)._webStandardTransport;
inner._initialized = true;
inner.sessionId = existingSessionId;

This works today (SDK 1.29.0) because _initialized and sessionId are plain TS private fields (runtime-writable JS properties), but it's fragile — any move to ES # private fields or internal restructuring silently breaks it.

Proposed API

A public method on both WebStandardStreamableHTTPServerTransport and StreamableHTTPServerTransport:

/**
 * Adopt an existing session ID on this transport, bypassing the
 * initialize handshake. Use this for multi-node deployments where
 * session state is externalized (e.g. to a database) and transports
 * are constructed per-request.
 *
 * @param sessionId - A previously-initialized session ID
 */
adoptSession(sessionId: string): void {
  this._initialized = true;
  this.sessionId = sessionId;
}

For the Node wrapper (StreamableHTTPServerTransport), it would delegate to this._webStandardTransport.adoptSession(sessionId).

Evidence

We built a POC that exercises 4 scenarios (fresh transport rejection, direct monkey-patch, synthetic-init workaround, Node wrapper field-reach). All 4 pass on SDK 1.29.0.
Our production workaround includes a runtime writability assertion that crashes loudly if _initialized becomes non-writable, pinned SDK version, and the POC as a pre-upgrade regression gate.

Context

• ADR documenting this decision: internal (ADR-017, MCP session externalization)
• The session state that needs to persist is minimal: just sessionId + "initialized" flag (~100 bytes)
• eventStore, _streamMapping, _requestResponseMap are all per-request transient and don't need rehydration
• This pattern is relevant to any multi-node or serverless deployment of an MCP server

Alternatives considered

1. Synthetic initialize — drive a fake initialize request through the transport. Works but costs 5-10ms and triggers McpServer side effects (capability negotiation, notifications).
2. EventStore-based resumption — the eventStore hook doesn't help because it replays events, not session state.
3. Direct private field mutation — what we're doing now, with the fragility noted above.

A public adoptSession() method would make option 3 safe and officially supported.