fix(server): return stdio parse errors

[he-yufeng]

Summary

• make the stdio server return JSON-RPC error frames when incoming lines fail JSON-RPC parsing/validation
• recover a simple request id from the raw line when possible so id-bearing malformed requests do not stay pending
• keep later valid stdio messages flowing after an invalid line

Fixes #2751.

To verify

• uv run pytest tests/server/test_stdio.py -q
• uv run ruff check src/mcp/server/stdio.py tests/server/test_stdio.py
• uv run pyright src/mcp/server/stdio.py tests/server/test_stdio.py
• git diff --check

[he-yufeng]

Updated the coverage for the parse-error path after the first CI run showed the full test suite passing but coverage dropping below 100%.

Local checks on the updated branch:

• uv run pytest tests/server/test_stdio.py -q (5 passed)
• uv run ruff check src/mcp/server/stdio.py tests/server/test_stdio.py
• uv run pyright src/mcp/server/stdio.py tests/server/test_stdio.py
• python -m py_compile src\mcp\server\stdio.py tests\server\test_stdio.py
• git diff --check

I also attempted the full coverage command locally on Windows/Python 3.13, but it fails in existing integration transport tests before coverage reporting (HTTPStatusError 502 / websocket closed-before-status-line cases). Those failures are outside this stdio parse-error change; the targeted stdio coverage has been expanded and pushed.

[he-yufeng]

Follow-up pushed in f76a0b3 after the remote coverage failure on stdio.py line 51.

I added the missing malformed-JSON numeric-id regression case so the fallback int(raw_request_id) branch is covered.

Local validation:

• uv run pytest tests/server/test_stdio.py -q: 5 passed
• uv run ruff check src/mcp/server/stdio.py tests/server/test_stdio.py: passed
• uv run pyright src/mcp/server/stdio.py tests/server/test_stdio.py: 0 errors
• python -m py_compile src\mcp\server\stdio.py tests\server\test_stdio.py: passed
• git diff --check: passed

[he-yufeng]

Follow-up pushed in b63f5f8 after the Python 3.14 lowest-direct job exposed an unrelated stderr exact-match fragility in the stdio interaction test.

The test still verifies that the subprocess clean-exit marker is passed through stderr, but now checks the marker as a suffix so dependency warnings emitted by the child interpreter do not fail the test before the server's own line.

Local validation:

• uv run pytest tests/server/test_stdio.py tests/interaction/transports/test_stdio.py -q: 7 passed
• uv run ruff check src/mcp/server/stdio.py tests/server/test_stdio.py tests/interaction/transports/test_stdio.py: passed
• uv run pyright src/mcp/server/stdio.py tests/server/test_stdio.py tests/interaction/transports/test_stdio.py: 0 errors
• python -m py_compile src\mcp\server\stdio.py tests\server\test_stdio.py tests\interaction\transports\test_stdio.py: passed
• git diff --check: passed

[StantonMatt]

Checked head b63f5f8 with the stdio paths this changes.

Local validation passed:

• uv run --frozen pytest tests/server/test_stdio.py tests/interaction/transports/test_stdio.py -q -> 7 passed
• uv run --frozen ruff check src/mcp/server/stdio.py tests/server/test_stdio.py tests/interaction/transports/test_stdio.py
• uv run --frozen pyright src/mcp/server/stdio.py tests/server/test_stdio.py tests/interaction/transports/test_stdio.py

I also ran a small stdin/stdout probe with a malformed id-bearing line followed by a valid ping: the malformed line produced a JSON-RPC parse error with id "bad-1", and the later valid request still arrived on the read stream. That covers the main regression boundary from #2751 for me.