fix: restore main-only v3.9 fixes that never reached dev

[gg582]

Brings back to dev the main-only work that v3.9 shipped but that never reached dev.

How this was found: I compared v3.9 against current dev file by file. For every line that exists only in v3.9, I checked whether dev had deliberately replaced it.

• Most such lines are older code that dev has since superseded: the old static task ring, the old async claim spin, the old gc.h docs, the HTTP/3 changes. Those are left alone.
• What remained is listed below. Each item was added on main only, mostly in Jiho Lee's 2026-09-14 batch, and before the histories were joined it was never ported to dev.

Commits

1. fix(api): restore the inert CWIST_ENDPOINT_PUBLIC_FIXED bit (79144d1)

   • v3.9's endpoint_opts.h defines CWIST_ENDPOINT_PUBLIC_FIXED (1u << 3).
   • When perf(bdr): per-connection cursor fast path for cached GET replies #194 removed the public FIXED cache, cf69362 decided to keep this bit "defined (inert) on both branches", but dev never received the definition. Code built against v3.9 that sets this bit does not compile against dev.
   • This commit takes v3.9's endpoint_opts.h, which also replaces dev's stale "proposed opt-in, draft PR docs: propose safe public FIXED response cache ADR #85" text.
   • make api-baseline output is unchanged, because the baseline does not track macros.

2. fix: complete shutdown writes and transfer queue ownership (7dfe8be, author Jiho Lee, port of 3122858)

   • kqueue.c: the BSD io_queue now disowns the queue and each submitted job from the allocating thread's full-GC sweep, as the other backend already does. A submit whose job allocation failed now returns an error instead of dereferencing NULL.
   • http.c: during shutdown, a coalesced batch is sent with a blocking write, instead of being parked on a reactor callback that will never run.
   • tests/test_gc_job_handoff.c is wired into TEST_TARGETS, running the gc, nogc and runtime modes, plus gc and nogc again under -DNDEBUG.
   • The io_queue.c and scheduler.c halves are already on dev from dcb561b.

3. fix(http): isolate blocking pipelined request bodies (dfb75b9, author Jiho Lee, port of c5f289e)

   • cwist_http_receive_request() passed everything read so far to the header parser. When a pipelined request was already in the buffer, its bytes were taken as the current request's body.
   • This is a real bug on dev. The restored check_blocking_pipeline cases fail on dev without the fix (body-size assertion) and pass with it.

4. docs(app): cwist_app_listen() defaults to one worker per core again (bd9ec85)

   • The code was reverted by perf(app): revert default workers to one per core (fixes webserver latency gate) #335, but the header comment still described perf(app): default worker processes to one per three cores #333's "one per three cores".

Deliberately not restored

• The bdr_ctx field comment "Legacy standalone BDR storage; never replayed by HTTP dispatch". Dispatch does replay it, through the perf(bdr): per-connection cursor fast path for cached GET replies #194 pinned-cursor path, on both branches, so dev's comment is the accurate one.
• Unused or deleted files: tests/support/cjson/cJSON.h (nothing includes it), tests/bench_gc_alloc_overhead.c (never wired), docs/roadmap-v3.5.md (not referenced), scripts/ci/rust_hello_smoke.sh (deleted on dev in 45f9f7a), and the legacy othello-web Go files (also deleted on main).

Verification

Under ASan + UBSan, these pass: test_gc_job_handoff, test_http_pipeline, test_shutdown, test_https, test_https_park, test_https_park_full_gc, test_http2, test_conn_registry, test_full_gc_sweep, test_io_queue_full_gc, test_metrics, test_websocket. check_test_wiring.py and cwist audit --gate are clean. kqueue.c only builds on BSD, so the kqueue CI job covers it.