type-context: fix async disconnect race condition found by antithesis

[ryanofsky]

Fix a race condition reported in #348 where if a disconnect happens during an IPC call that uses a worker thread (an IPC call taking an mp.Context parameter), it can trigger a read-write race detected by TSAN, and also theoretically cause a null pointer dereference (described in the commit message).

The race condition happens because when RpcCallContext::getResults() is called for the first time, it checks the connection state. So currently if there is a disconnect, when the worker thread calls getResults, this can read connection state at the same time capnproto writes as it processes the disconnect.

Fix this issue by calling getResults once from the event loop thread before executing the IPC call on the worker thread, so the results message pointer will be cached, and future calls to getResults from the worker thread won't access the connection state or have any race condition.

This is a one-line fix with many comments and a test.

[DrahtBot]

The following sections might be updated with supplementary metadata relevant to reviewers and maintainers.

Reviews

See the guideline and AI policy for information on the review process.

Type	Reviewers
ACK	xyzconstant

If your review is incorrectly listed, please copy-paste <!--meta-tag:bot-skip--> into the comment that the bot should ignore.

Conflicts

Reviewers, this pull request conflicts with the following ones:

• #342 (Allow request cancellation for wrapped C++ methods by xyzconstant)
• #336 (proxy-io: Reference-count Connection objects by ryanofsky)

If you consider this pull request important, please also help to review the conflicting pull requests. Ideally, start with the one that should be merged first.

[xyzconstant]

Code review ACK 914dc83.

This one-line change builds an empty Results struct for each context-aware server method during request setup, on the event loop thread before the wrapped method runs on the worker. The first getResults call caches the response body, so subsequent calls from the worker return the cache rather than reading the connection state.

LGTM.