runtime: the handshake failure paths leak an outbound writer task (residual from C-01/C-02/C-04) #287

Found by the adversarial seat during the pre-merge code review of #286, and reproduced by it — filed as a residual rather than fixed there because it is bounded and cannot redeliver, so it did not warrant holding that merge.

What

#286 added a bounded outbound writer task (_drain_outbound) and wired its teardown into two paths:

• _cleanup_client — cancels outbound_writer and nulls both managed.outbound and managed.outbound_writer
• disconnect_server — cancels the writer, then sweeps background tasks

There is a third teardown path that was not in that plan's acceptance list: handshake failure. The except-paths in _connect_stdio (around manager.py:2456) and _connect_remote_stream (around manager.py:2780) cancel only read_task and stderr_task. If a writer was started before the handshake failed, it survives that pop.

The reviewer reproduced a live _drain_outbound task after the except-path pop.

Severity — bounded, stated accurately

• Retry-capped. MAX_CONNECTION_RETRIES = 3, and the reconnect loop makes 3 attempts, so this is at most a small number of leaked tasks per reconnect cycle rather than unbounded growth.
• Cannot redeliver onto a new process. The leaked writer holds the discarded ManagedClient. All three constructors (_connect_stdio:2413, _connect_remote_stream:2731, adopt_process:3641) allocate a fresh object, so no production path reuses one.
• Net effect is a slow task/queue leak against a server whose handshake repeatedly fails — not a protocol-correctness defect at the downstream boundary.

Suggested fix

Include outbound_writer in the handshake except-paths the same way #286 included it in _cleanup_client, and null outbound/outbound_writer there too so the postcondition holds on every teardown path rather than on two of three.

Prove it with a measured mutation, per the convention these runtime fixes have used: assert no live _drain_outbound task survives a failed handshake, then remove the fix and confirm the test is red for a named reason with the mutation confirmed applied via git diff --stat. Name the test so it is selected by the existing filter (it must contain one of c01_, c02_, c04_, stalled_sink, outbound_queue, outbound_writer) — a name matching no filter runs as part of nothing and still exits 0.

Two other non-blocking nits from the same review

• disconnect_server cancels the writer but does not null managed.outbound. Harmless today because the object is discarded, but the same postcondition argument applies.
• A frame whose "method" is present but not a string falls through to the pending-response lookup. It is not a valid JSON-RPC request, so this is a robustness nit at an untrusted boundary rather than a correctness bug.

See #232, #286.