fix: drain accepted JSON-RPC requests after read EOF

[matthewchen94]

Fixes #2678.

This changes the JSON-RPC dispatcher shutdown path so read EOF does not immediately cancel request handlers that were already accepted. The dispatcher now:

• marks the connection closed and wakes outbound waiters after read EOF,
• keeps the write stream open briefly for accepted inbound requests to finish their response writes,
• bounds that drain window to avoid hanging shutdown,
• tracks active inbound requests through the response write window, not just while they are present in _in_flight.

This targets stdio servers driven with redirected stdin, where EOF can arrive immediately after the last JSON-RPC request is read while an async tool call is still mid-await. In that case the response should be flushed to stdout before shutdown.

Validation:

• python3 -m py_compile src/mcp/shared/jsonrpc_dispatcher.py tests/shared/test_jsonrpc_dispatcher.py passed on the patched files.
• Added a dispatcher-level regression test that accepts a request, closes the read side, and asserts the response is still written.

[cubic-dev-ai]

1 issue found across 2 files

<sub>Reply with feedback, questions, or to request a fix.

Fix all with cubic | Re-trigger cubic</sub>

[matthewchen94]

Updated this PR to address the review and failed CI:\n\n- Scoped EOF draining behind an explicit dispatcher opt-in so default transport-close semantics still cancel in-flight handlers immediately.\n- Enabled that opt-in only for stdio read streams, where redirected stdin can EOF after the final accepted request while stdout still needs to flush the response.\n- Tracked the transport_builder rejection path in the drain counter, so spawned INTERNAL_ERROR responses are not cancelled before reaching the peer.\n- Merged latest main and added regression coverage for both accepted handler responses and builder-rejection responses.\n\nAll CI checks are green on f4a8fc9.

[matthewchen94]

Updated this PR onto latest main with a merge commit, without force-pushing.

What changed in the refresh:

• rebuilt the stdio EOF-drain change on current main
• preserved the new runner cache-hint handling that landed after the earlier PR head
• kept EOF draining scoped to stdio/read streams via drain_inbound_on_read_eof
• propagated the opt-in through both current dispatcher construction paths
• kept regression coverage for both accepted handler responses and transport-builder rejection responses

Local validation on the rebuilt files:

python3 -m py_compile \
  src/mcp/shared/jsonrpc_dispatcher.py \
  src/mcp/server/runner.py \
  src/mcp/server/stdio.py \
  src/mcp/shared/_context_streams.py \
  tests/shared/test_jsonrpc_dispatcher.py

The PR is mergeable again; CI is rerunning on 14ed0cb.

[matthewchen94]

Quick maintainer ping on this one.

The PR is now mergeable and the refreshed head 14ed0cb has the full required matrix green:

• pre-commit
• server/client conformance
• Python 3.10–3.14 test matrix on Ubuntu/Windows
• docs/readme snippets
• all-green

The refresh was done as a merge commit on top of the previous PR head, not a force-push, and it preserves the newer runner cache-hint changes from current main.

I don't plan to change this further unless there is specific review feedback, because the current patch is scoped to the stdio EOF-drain behavior and CI is clean.