perf(reactor): bound epoll/kqueue rounds with the drain-chunk cadence

[gg582]

Summary

• What: ports the existing cooperative-queuing drain cadence (CWIST_REACTOR_DRAIN_CHUNK) from the io_uring dispatch loop to the epoll and kqueue loops in cwist_reactor_run().
• Why: the epoll loop (now the default backend under seccomp/no-uring via the RX-epoll fast path) drained foreign-thread posts (cwist_reactor_post) only once at the top of each round, then ran the full epoll_wait array (up to 1024 events) with no drain point in between. A background/foreign-thread completion arriving mid-round waited behind every remaining event's handler, so its tail latency scaled with the whole same-wake batch — the exact P99.999 mechanism issue [BUG][Plan/CWIST 3.5] Enhance bad P99.999 latency #25 fixed for io_uring, un-fixed here.

A/B result (meaningful ✅)

Deterministic harness bench_cooperative_queuing_epoll (512 pipes readable in one round × 80 µs fake handler each, one foreign-thread post mid-batch; baseline binary built from HEAD, 5 interleaved rounds × 10 runs, re-verified locally):

epoll backend	post-to-callback latency
baseline (HEAD, no mid-round drain)	mean 41.6–42.2 ms, max 46.6 ms
CWIST_REACTOR_DRAIN_CHUNK=8 (profile default)	mean 0.43–0.49 ms

≈ 90–99× tail improvement, matching the theoretical bound (chunk × handler cost). chunk=0 reproduces baseline exactly (legacy opt-out preserved).

No regressions: io_uring-backend bench unchanged (that loop is untouched); cwist_bench + wrk -t8 -c256 on the forced-epoll plain request path shows identical throughput/latency across legs (no foreign-thread posts in that workload, and the added check short-circuits on drain_chunk == 0).

Changes

• src/sys/io/reactor.c: epoll + kqueue loops drain posts every N dispatched callbacks (i + 1 < n guard mirrors the io_uring head != tail guard); reactor_drain_chunk() hoisted out of the __linux__ guard.
• tests/bench_cooperative_queuing.c + Makefile: new bench_cooperative_queuing_epoll target (-DCWIST_BENCH_EPOLL pins CWIST_REACTOR_BACKEND=epoll); the io_uring twin now pins its backend symmetrically.
• .github/workflows/perf-epoll-drain.yml: path-filtered A/B scenario — same binary, CWIST_REACTOR_DRAIN_CHUNK=0 vs 8, 5 interleaved rounds on push to dev (+ io_uring no-regression leg), artifacts uploaded.
• Docs (en + ko): CWIST_REACTOR_DRAIN_CHUNK and the cooperative-queuing page now state the cadence applies on all three backends.

Test plan

• make bench_cooperative_queuing_epoll A/B rounds (numbers above)
• test_reactor_drain_chunk, test_reactor_wake, test_reactor_timer (default + forced epoll), test_rx_epoll, test_http_fairness, test_async_defer — pass
• scripts/ci/check_test_wiring.py, scripts/ci/check_env_knobs.py — pass
• New perf-epoll-drain.yml workflow runs on this PR's merge to dev (path-filtered to the workflow file itself)