MDK2 continuous early GC repair
The activation repair starts the installed observer before ordinary command65, at the existing personally reviewed startup3584 main-menu observe16. It retains that same observer through ordinary New Game and Medium controls. Installation was already early; the diagnosis is f441255216c45f00f6be5ac6f201fe474144e5a2. The terminal-report remedy from published3dafc52d0 and the missing-entry rejection from original evidence3c0035701 remain intact.
Admission requires the reviewed main-menu visual classification, original native New Game readiness, frozen16ms, exact3584 step count, PC47d9c8 and the reviewed saved-return chain reaching402176 before conditional PUSH402186/CALL402187. These stack words admit the checkpoint; they never become observer records. Earlier main-menu observations remain dormant. A later unsupported first main-menu attachment is rejected.
The driver correlates the reviewed scene, actual native-selector reservation, trusted Enter down/up receipts, exact input-history prefix, queue suffix, active input, ACK and owner/epoch fences. New Game must reach native full rest after release before a fresh difficulty review can authorize Medium. Medium must likewise complete its press/release/full-rest phase before later construction scenes are admitted. Each quantum checkpoints and resumes the same observer; neither records, counters nor the fixed whole-sequence deadline are reset.
A native negative exposed a separate containment defect: replacing an owning instance made the binding terminal, but a subsequently queued slice could execute unobserved. The run wrapper now returns before execution after such fatal binding/continuation failure. The first terminal report remains queryable and unchanged. Existing real import-failure behavior stays covered by the integration suite.
The added native regression extends the shipped real-Worker nativeOne/workerMain fixture for both accepted modules. An explicit synthetic minimal Lua state and interned string heap are derived from original PE initialization and collector instructions. The fixture executes unchanged402176, actual PUSH/CALL443070, the complete original collector and its real RET to40218c, then stops at the ordinary continuation callee452450. It splits only at queued slice/checkpoint boundaries. Successful observed, late-prefix and unobserved control paths perform no register or memory edits after initialization, use no collector stub and invent no observer records. The original root traversal marks an initially unmarked string and the original sweep clears that mark. Genuine entry is boundary3 and matching return is boundary1023 in both modules. A pending entry survives a frozen review and resume; observed registers, complete shared memory and commands match the identical unobserved control. Native counts are monotonic and the original absolute deadline is retained.
Late start at actual449096 still rejects return40218c for missing entry, preserves the complete first report and executes nothing on the next queued slice. Both accepted modules cover changed instruction pins, replacement owner, changed epoch, changed ESP, wrong caller, expired review and lower test-only boundary/record limits. The production limits are unchanged.
A separate explicitly synthetic producer fixture uses the actual retained selector, Spool, inspection guard and driver. Its sequence is main-menu observe16, freshly reviewed New Game press16/release32/fullrest, fresh difficulty, then Medium press16/release32/fullrest: one start, four resumes, five checkpoints. Unsupported stacks/PC, non-rest native state, stale reviews, bad ACK, untrusted or foreign input, changed input history and expiry reject before dependent execution. Its synthetic readbacks qualify protocol composition, not original gameplay.
Validation passed: all178 integration tests (integration-final.log), all17 focused early-GC/resolver tests (rooted-collector-focused.log),45 runtime contracts plus the observer suite with native parity for both modules (runtime-observer-focused.log), and full flock /tmp/wine-build.lock bash tools/build.sh (build.log). The first full run passed177/178 and exposed a stale resolver-test message assertion (accepted/current); the assertion now matches the already shipped current-binding contract, and the final rerun is green. Both rebuilt modules retain hashes e620f3adfc2850580004e204bc1fea495844e942806866f8ada24c215212e412 and 66b37bd76da5a7b49c063b4eb0d44c898aaf00ff41e000a6fbd87c01edc136c8. The contracts remain35 pins,502 compared bytes and534 address translations per boundary,50M boundaries/360 seconds/512 records, frozen16ms,512/544 menu limits and113 commands. No timer/render/load forcing, synthetic readiness in production, cap raise, shared-helper change or original game/browser run is part of this task. These genuine but synthetic Worker tests do not establish that the original reaches Medium within these caps.
Evidence is retained in scratch/runs/20261105-mdk2-continuous-early-gc-repair/: exact Git SHA, relevant binary diff and new-note patch; run-specific JSON and logs; existing PE/media paths and SHA-256. Git is the source archive. No source tree snapshots, copied modules or copied media were made. BOTH Stronghold0130 and ReturnFire0150 FINAL runtime/file releases ALLBUILDSFINISHED preceded source mutation and module execution; dependencies-cleanup-independent.json independently verifies their normal exact cleanup. Box ID/state/expiry: NA; coordinator owns shutdown.