fix(client): abort-on-close watcher busy-spins while events are pending

[krynju]

Repurposed — the original readbytes!→readavailable change is dropped: on a real HTTP.Stream, readbytes! returns per-chunk (verified empirically), so the original premise was wrong.

Problem

The abort-on-close watcher introduced in #98:

while isopen(stream_to)
    wait(stream_to)
    yield()
end

wait(::Channel) returns as soon as data is available, not when the channel closes. So whenever an event has been put! but not yet consumed — e.g. the consumer is busy converting the previous event — this loop spins hot. Measured live on a k8s watch stream: 100% of a core in system time (wchan=0, running thread, 400 stime ticks per 4s) for as long as any event sat in the buffered channel.

Fix

Poll isopen with a 250ms sleep. The watcher exists only to abort the blocked read when the consumer closes the channel; 250ms of extra abort latency is irrelevant, and the loop now costs nothing while idle.

Found while diagnosing deaf k8s watches in JuliaHub's JobLoops (root cause was in Kuber.jl — see JuliaComputing/Kuber.jl#67 — but this spin was compounding it by starving the same thread pool).

Bumps to 0.2.6.

🤖 Generated with Claude Code