[gc] Graduate incremental collection: budgeted scanner coverage, assists for all phases, default-on (successor to #6083)

[proggeramlug]

Problem

Automatic collections in compiled programs are synchronous, unbounded, alloc-point STW cycles because the budgeted (incremental) machinery is structurally unreachable:

1. ~50 of ~57 registered root scanners have no budgeted step form, and any unbudgeted scanner blocks the sliced stepper: predicate gc/roots.rs:286-307, registrations gc/mod.rs:289-454.
2. The alloc-point direct arm (gc/policy.rs:1232-1284) forces the conservative scan (:1254), which makes the copying fast path ineligible (gc/copying.rs:769-774) — so the collection lands in GcCycleState::new_minor_fallback(...).run_to_completion() with an unbounded budget (gc/cycle.rs:1002-1009).
3. PERRY_GC_INCREMENTAL=1 exists but is off by default and has a runaway hole: mutator assists only advance 4 phases (gc/cycle.rs:47-55, gc/policy.rs:1772-1778), so a compute loop that never reaches the event pump parks a cycle at the sweep boundary indefinitely with the incremental-mark barrier still enabled (gc/cycle.rs:1041-1044) → unbounded RSS.

Proposed fix (staged)

1. Convert the ~15 highest-traffic scanners to budgeted form (runtime handles, promise, timer, class/symbol tables already are; object caches, descriptor tables, closure caches, error props, fs/process tables are not).
2. Let mutator assists drive all resumable phases (sweep is already sliced), or add a debt cap that escalates to a synchronous finish — GcDebtSnapshot (gc/telemetry.rs:446-468) already measures exactly this but is telemetry-only.
3. Graduate PERRY_GC_INCREMENTAL to default-on once 1+2 land.
4. Route alloc-point trigger crossings to the pump-boundary moving minor via the existing deferral flag (gc/policy.rs:1246-1251) by default once loop-poll coverage lands.

Context

2026-07-09 GC audit. Successor to #6083, whose closure shipped the trigger re-arm fix and the off-by-default incremental mode but left the default path fully STW.

[proggeramlug]

Stage 1 (close the sweep-parking hole) is up as #6216. Root cause + fix: allocation-side mutator assists were gated to the four mark phases only (mutator_assist_honors_budget), so under PERRY_GC_INCREMENTAL a pure compute loop that never pumps the event loop would advance a budgeted cycle to AtomicFinalize and park there forever — mark barrier stuck on, nothing swept, unbounded RSS. #6216 lets assists drive every resumable phase (finalize/sweep/reclaim are already budget-sliced and reentrancy-from-alloc is proven by the existing direct-fallback mark-sweep), so the cycle now completes from allocations alone and RSS stays bounded.

Deliberately out of scope here (tracked follow-ups, both belong to the eventual default-on flip, not this correctness fix):

1. Slice AtomicFinalizeSubphase::WeakProcessing (currently one unbudgeted arena_walk_objects → a single O(live-weak-holders) spike per cycle when an assist lands on it). Pause-quality, not correctness.
2. Stage 2: enumerate remaining incremental correctness gaps and flip PERRY_GC_INCREMENTAL default-on.

This issue stays open to track Stage 2.

[proggeramlug]

Stage-2 progress from the measured A/B campaign (#6224 → PR #6226): debt-proportional assist pacing landed together with four budgeted-cycle soundness fixes the working pacing exposed (minor-cycle mark barrier w/ nursery-scoped shading, final seed drains at both barrier-off points, whole-cycle allocate-black for runtime-path allocations, and synchronous collections draining a parked budgeted cycle first). Differential stress: 7/7 exact-checksum clean incl. manual-gc + force-evacuate; was 7/7 SIGSEGV / silent 2,890-node loss.

Measured trade-off with everything active (N=3 medians): worst pause 636→118 ms (5.4×), p999 488→111 ms, throughput parity, RSS 1.8–2× STW (was 6–22× unbounded). p99 7→53 ms is the pacer honestly paying for collection.

Remaining before a default-on decision: (1) the stack-local remark phase (only-in-a-register/local pointers with no heap store during the cycle window) — now the single known soundness gap; (2) #6179 classifier to shrink the ~2× RSS overhead (exact valid-ptr BTreeSet held across sliced cycles + allocate-black floating garbage); (3) WeakProcessing slicing for the residual pause spike.