Optimizer: execute rooted class definition effects with exact activation recovery

Parent: #467. Executed on main3100f072, ReleaseSafe/Darwin, real collector.

Three required-bytecode functions returning plain, derived, and private-field classes are warmed64 times. Node and zig-js both produce[7,8,9]. Every function publishes a ready optimizer artifact, but its only runtime descriptor is at origin0 (enter_block) and its effect side-exit is also at origin0 with an empty operand stack. No class construction or heritage operation appears in those artifacts. Artifact presence is therefore not native class-effect admission.

Implement ordered class lexical-environment entry/exit, heritage's exact two-result transport, per-evaluation PrivateEnvironment preparation, and canonical class template evaluation, including computed/inferred names, captured activation views, static blocks/fields, methods/private identities and constructor initialization. Preserve precise rooted input/result transport, active environment/private/strict state, exact handler/unwind and completed-effect recovery, and callback budget/termination/OOM behavior. Unsupported surrounding control must still recover exactly while these owned effects actually execute.

Acceptance: executed native descriptor/artifact witnesses for each effect; forced off/on and tree/required-bytecode oracles in serialized/no-GIL Debug/ReleaseSafe/unsuppressed TSan; actual moving-GC and exact exception/no-replay/step/fault recovery; exact-parent class/super/private/eval failing paths; full unchanged fuzz and threading gates. Do not infer completion from published artifacts or working fallback. General class/tier parent acceptance remains separate.

Progress on main:\n\n- 40133162 feat(jit): execute class definition effects natively\n - native ordered enter_block/exit_block, prepare_class_heritage (two results), prepare_class_private_environment, and eval_class dispatch\n - optimizer graph admission, descriptor/result-count verification, scratch/result transport, and rooted class inputs\n - executed descriptor and runtime-operation witnesses in src/vm.zig\n\n- 98d4f9c test(jit): cover computed class recovery\n - computed class names, computed methods/static fields, private field, heritage, and static block\n - throwing static block increments exactly once per call through optimizer recovery (no replay)\n\nFocused serialized results (Zig 0.18.0-dev.146, aarch64):\n- optimizer executes native class definition effects: 2/2 passed\n- optimizer preserves computed class names and static blocks: 2/2 passed\n- native class exception resumes without replaying static effects: 2/2 passed\n\nIssue remains open pending moving-GC, OOM/step/fault recovery, forced optimizer-off/on exact-parent differential, and serialized Debug/ReleaseSafe/unsuppressed TSan/fuzz/relevant conformance gates.

Additional evidence on main:\n\n- cbdcd58 test(jit): witness native class moving gc\n - warms the existing computed-private class builder through optimizer publication\n - asserts an executed eval_class descriptor and runtime-operation witness\n - forces the existing moving collection path and preserves the relocated subject/result\n\nSerialized focused result (Zig 0.18.0-dev.146, aarch64):\n- computed class private environment survives an actual moving collection: 2/2 passed

JIT gate result after eff50cd:\n\n- zig build test-jit: 161 passed; 1 skipped; 0 failed; 0 leaked.\n- The only initial failures were stale side-exit matrix entries for class effects; eff50cd removes those now-continuable class operations from the terminal-side-exit fixtures.\n- No concurrent corpus or unit jobs were run.

Exact-parent differential added and pushed in 453ad44:\n\n- vm: class definitions agree across tree bytecode and native parents\n- tree-walker: passed\n- required bytecode with JIT disabled: passed\n- required bytecode with JIT enabled: passed, including an executed eval_class descriptor witness\n\nFocused serialized result: 2/2 passed.

Thread corpus gate completed on Zig 0.18.0-dev.146: zig build threads-test passed 246/246 corpus files. This includes moving-GC, OOM, termination, stack limits, JIT/OSR, atomics, synchronization, and VM-state cases; no failures were reported. The issue remains open pending no-GIL, bounded fuzz, sanitizer, and remaining class-specific recovery evidence.

Bounded fuzz gate completed: zig build threadfuzz -Dfuzz-iters=400 passed 400 deterministic programs from seed 1 with 0 failures in 209,008 ms. The issue remains open pending no-GIL, sanitizer, and remaining class-specific recovery evidence.

No-GIL gate completed on Zig 0.18.0-dev.146: zig build threads-test -Dthreads-parallel-js=true passed 253/253 corpus files. This included no-GIL moving-GC, OOM, termination, JIT/OSR, optimizer invalidation/reclamation, atomics/races, and VM-state coverage with no failures. The issue remains open pending sanitizer evidence and remaining class-specific recovery checks.

Pushed ff795be4 (test(jit): cover class step recovery). The focused VM test passed 2/2: a native eval_class witness arms the step budget from a static-block host callback, the call exits with RangeError, the callback runs once, and static effects remain exact-once (24 warm-up effects, 25 after the interrupted evaluation). No source replay occurred. The issue remains open pending full JIT rerun, sanitizer evidence, and final class recovery audit.

Full JIT gate after ff795be4: zig build test-jit passed 161 tests, skipped 1 platform-specific Linux writer test, with 0 failures and 0 leaks. The pushed class step-recovery coverage is included. Proceeding to the narrower unsuppressed TSan JIT gate before considering the full suite.

Unsuppressed TSan JIT gate completed: zig build test-jit -Dtsan=true passed 161 tests, skipped 1 platform-specific test, with 0 failures and 0 leaks. No sanitizer report was emitted. Debug and TSan JIT evidence are now green; ReleaseSafe JIT is next.

ReleaseSafe JIT gate completed: zig build test-jit -Doptimize=ReleaseSafe passed 161 tests, skipped 1 platform-specific test, with 0 failures and 0 leaks. Across Debug, ReleaseSafe, and unsuppressed TSan, the JIT gate is green. I am adding one native class OOM rollback witness before the final full-suite decision.

Pushed 6f105d3c (test(jit): cover class OOM recovery). The focused VM test passed 2/2: native eval_class receives an injected OOM during a static block, preserves one completed effect without replaying it, returns cleanly to the caller environment with depth 0 and empty GC root stacks, and records native eval_class execution. This closes the class-specific OOM/rollback coverage gap.

Final pushed-tree verification for #1110 is complete.

Implementation and recovery coverage landed directly on main in 40133162, 98d4f9c, cbdcd58, eff50cd, 453ad44, ff795be, and 6f105d3.

Final exact-tree gates:

• Debug JIT: 161 passed, 1 skipped, 0 failed, 0 leaked
• ReleaseSafe JIT: 161 passed, 1 skipped, 0 failed, 0 leaked
• Unsuppressed TSan JIT: 161 passed, 1 skipped, 0 failed, 0 leaked; no sanitizer output
• GIL thread corpus: 246/246
• no-GIL thread corpus: 253/253
• deterministic thread fuzz: 400/400

Focused witnesses cover executed native class artifacts, exact-parent tree/bytecode differentials, moving GC, exception/no-replay, step-budget recovery, and OOM rollback/teardown. The lone skip is the existing platform-specific perf-jitdump case. All commits use glennmichaeltorregosa gtorregosa@gmail.com with no trailers or coauthors.