Re-land the fix and regression test for vuln-187

[harrishancock]

A draining read delivers its result to pumpToImpl through a kj event, and runs the stream's pull() callback along the way. A pull() that aborts the request's signal makes the canceler destroy the pump's coroutine frame from inside that event, so the event is destroyed while still firing and kj aborts the process with "Promise callback destroyed itself".

Neither isolate-lock round trip in the pump is awaited directly any more. Each is handed to the IoContext's task set, and the pump awaits an unrelated fulfiller, so dropping the pump destroys only that await and leaves the firing event to be unwound by its new owner.

Those tasks outlive the coroutine frame, so they reach the DrainingReader through a kj::Weak rather than the frame-owned kj::Own. Without that, the read lambda would run against a freed reader whenever the pump is dropped before the isolate lock is acquired.

The fix originally landed as 00e86e1 and was swept away by 4447562, a wholesale revert of the streams-cleanup-investigation merge. That first version captured the reader by reference, which is what the kj::Weak above replaces; this follows the later 2de66f75e, which had corrected it. Both are James's, as is the third attempt, f3f408768, which drops the code change and keeps only the test — not an option here, since the abort still reproduces without it.

The surrounding tree has moved on twice over: the draining read is no longer autogated, so the test needs no gate, and DrainingReader is a kj::PtrTarget, so the weak pointer comes from addWeakToThis() rather than a hand-rolled cell.

Assisted-by: OpenCode:claude-opus-5

[ask-bonk]

LGTM

github run