gc: test_gap_net_connect_bound_value flakes ~50% — uncaught throw in the native-module namespace read, collection-window class

[proggeramlug]

test_gap_net_connect_bound_value fails intermittently (~50–60% of runs) on main-gate sweeps: Perry exits 1 after printing only the first output line — the uncaught throw happens between console.log("connect:", typeof net.connect) and console.log("createConnection:", typeof net.createConnection), i.e. in the native-module namespace property-read path, before any socket/pump activity.

Evidence:

• Same head SHA flips red/green (92f036a7b: red 07:03, green 07:25, red 09:17; d1d992966: 4/5 red) → a collection-window bug, not deterministic state.
• Present since the earliest completed sweeps (08-18 13:03, 7cadb938e) — the introduction window is anything ≤08-18, sweeps simply didn't run before.
• Not fix(ext-net): give the pump's in-flight dispatch GC custody (#8259) #8265's dispatch-custody change: that PR only touched js_ext_net_drain_pending's dispatch arms and scan_net_roots; the crash site executes before any listener registration or pump tick.
• Suspect family: the recent native-module namespace / dynamic-object-get work — fix(gc): root raw pointers in dynamic-object get, add a PROCESS_EMITTER scanner #8294 ("root raw pointers in dynamic-object get"), fix(runtime): enumerate native module namespace values #8269 (native-module enumeration), fix(cjs-wrap): stop HIR from dropping built-in require bindings in wrapped modules #8343 (cjs-wrap require bindings; this test reaches net via createRequire).

Intermittent = an unrooted register/window per the rooting playbook; hunt with the seeded instruments (PERRY_GC_SCHEDULE_RATE=1 PERRY_GC_SCHEDULE_SEED=N PERRY_GC_PROTECT_FROMSPACE=1) on this one test — the protect arm should turn the flake into a deterministic fault at the real site.