fix(jsc): wake passive collectors for synchronous GC on main

[steipete]

Main-branch companion of #21. The collector fix and native probe are identical; this branch uses its existing qualification script and inline receipt gate.

A worker can stall inside synchronous inspector GC while the main thread waits for its database work. A collector that has already entered passive draining checks mutatorWaitingBit only on entry. If its mutator starts waiting afterward and shared helpers are occupied by another heap, neither side wakes the collector to take over the marking work.

Notify the marking condition under its mutex after releasing the thread lock, and recheck the waiting bit inside the passive loop. The collector then unlocks and drains actively. The 14 added production lines close both the missing notification and the missing predicate check; collection policy, helper counts, heap limits, and application concurrency remain unchanged.

The native regression uses actual Heap operations and ParkingLot observations, holds the shared helpers, and requires collection to finish before releasing them. The unchanged engine positively exhibits both parked waiters, with no timeout; the fixed engine completes and preserves rooted objects. The regression is included in the existing native qualification and its receipt verifier.

Validation so far:

• macOS native control fails on the observed wait cycle; candidate passes. The integrated regression also passes on this PR's current main-branch base.
• 345 surrounding Bun Worker/inspector tests and 25 OpenClaw worker-pool, memory-reclamation, and cleanup cases pass with the isolated fix.
• Independent review of the complete change is clean at P2.
• A separately instrumented maintained OpenClaw 100-client / 50-active run reproduced 20 reply deadlines. One SQLite-worker GC overlapped 99.96% of a 35.2-second held database admission, with 49 writes queued behind it. Native stacks show the mutator waiting for collection, a passive collector, and occupied shared helpers. These diagnostics identify the mechanism; debugger-overlapping latency spans are excluded from attribution.

The corrected native fixture passes on both macOS and Linux. It releases its context inside the existing JS lock scope; the earlier Linux teardown abort is retained as invalid fixture evidence, not an expected control failure. The fresh exact-source Linux pair positively detects the unchanged-engine wait cycle without timing out (4.413 ms); the candidate completes before helper release (5.456 ms). Both preserve rooted objects.

The completed maintained OpenClaw 100-client / 50-active / 4-driver prime now passes with the fixed engine: 101 replies (51 during the window, 50 during drain), zero errors, ACK p50/p95 9.628/14.025 s, clean 2.567 s shutdown, and no worker-timeout text. Node 26.10.0 on the same host also passes with 103 replies, zero errors, and ACK p50/p95 10.386/14.821 s. First-wave ACKs arrive 7.511–9.630 s after the first send on the fixed engine and 6.626–8.647 s on Node.

These unscored primes use identical frozen OpenClaw source 4998572a3abb0e4ae1a156fe38ffd8ebb5cf8124 plus the separate cheap-counter patch in openclaw/openclaw#166104. Gateway CPUs are 0–3 on a Hetzner ccx53, with controllers/provider on 4–31; 60-second admission, 180-second reply deadline and 1,000 ms chunk pacing are unchanged. Both Bun arms were built from 667c4ab22cbf6b101b3376c550b81cabc8c00518 and engine f1e1ca1156c8cb3b468bec0e1989fbfa08899661, with identical release/non-LTO flags and only this 14-line runtime patch differing. This isolates the runtime repair on the original engine pin; the current PR bases additionally pass their own native builds and exact-head CI.

The rebuilt unchanged-engine control failed warmup before load, so it supplies no matched load-window latency estimate. Earlier published 667c and 670b baselines both failed 100/50 while Node passed; an earlier host-interrupted paired unit is excluded entirely. The exact original single 63-second ACK wave and three literal timeout messages were not reproduced. The native regression proves the missed-wakeup defect; Gateway profiling and successful fixed primes support its workload relevance without assigning every historical stall to it.

No engine merge or artifact publication is performed by this work. Upstream searches found no matching repair; WebKit#72377 is already represented in the base and addresses a different notification path.

The final same-shape passive-timeout audit also passes: 102 replies, zero errors, 20 observed pools with no timeout events, and clean 2.195 s shutdown. It adds no timers, worker preloads, heap reads, or retries; the uninstrumented prime above remains the latency comparison.