MDK2: why startup3584 did not reproduce pre-GC admission

Published original report cd4b2e74174231b62ab3aa0affc2d1b1bb077efd stopped correctly: command65 reached the exact-PC assertion in createTraceDriver.begin() with native thread1 at 0x4641c7, rather than 0x47d9c8. The earlier implementation treated a checkpoint observed in one original as reproducible at a fixed host-step count. The new retained original disproves that reproducibility assumption. Its stack also fails every one of the seven required saved-return positions. Accepting this PC would require a new native safety argument; changing the literal alone is unjustified.

The immediate cause is established. The exact native cause of the divergent execution paths, and a safe ordinary route from the new state to pre-GC observation, remain unresolved by retained evidence. No runtime change or retry was made for this diagnosis.

Inputs are the continuous original report, the earlier attachment analysis, and the actual retained runs scratch/runs/20261105-mdk2-continuous-original-medium/ (NEW) and scratch/runs/20261105-mdk2-post-medium-timer-onward/ (OLD). Diagnostic evidence is in scratch/runs/20261105-mdk2-reachable-checkpoint-cause/ (RUN). RUN/audit.py reads actual retained JSON, checks identities and hashes, and records its inputs in audit.json. It does not import or execute the game, driver, Worker or module. Git is the source archive; RUN records the audit Git SHA, relevant binary diff and new-file patches. Existing executable/media/module paths and hashes remain references, without copies.

Recorded controls and native states

Removing only timestamps, OLD and NEW browser/commands.jsonl agree exactly through command65. The startup prefix contains 56 ordinary step64 commands and eight startup captures including startup0 and startup3584. They reach3584 steps before command63 captures startup3584, command64 reads native selector state, and command65 requests the ordinary observe16. There is no New Game input in that prefix. Both startup3584 native selections have startscreen, selected0, command62 phase/raw/previous/active0, empty input queue, null active input and mouse mask0. These predicates support menu readiness, not a native call history.

Frozen steps OLD actual PC / ESP NEW actual PC / ESP
0 4800a5 / 074ffffc 4800a5 / 074ffffc
512 4515cd / 074ffd40 4515cd / 074ffd40
1024 4515cd / 074ffdb4 4515cd / 074ffb2c
1536 4515cd / 074ffd3c 4515cd / 074ffd3c
2048 45bd9e / 074ffb88 443494 / 074ffd24
2560 4515cd / 074ff8b4 4515cd / 074ff8b4
3072 44fc40 / 074ffd0c 47cef1 / 074ffc88
3584 47d9c8 / 074ffcd8 4641c7 / 074ffbe8

These are startup-N-native-main.json readbacks, not PCs inferred from pixels or the outer step result's eip:0x00000000. Divergence is already visible in ESP at1024; the sampled states sometimes coincide again. They cannot determine which run is globally ahead. Both advance512×16ms between captures, but OLD guestMs exceeds NEW by88ms at every sampled checkpoint, including startup0 (924 versus836). Thus freezing16ms does not equate startup clock epochs either. This offset is observed; its causal contribution is unknown.

NEW's bounded-trace-ownership-before.json and -after.json are identical at 3584. reply-65.json preserves the unsupported-checkpoint assertion at 03:04:29.676Z. The matching terminal slot0 report binds module e620f3adfc2850580004e204bc1fea495844e942806866f8ada24c215212e412 and says started:false, traced:false, observer:null, traceEnd:null. Its last native slice had budget6000, ordinal3323, completed=true, ranBlocks6000. This is one completed slice, not cumulative native boundaries or a history of budgets. The assertion precedes Worker invoke('start'), its initial sample and the dependent WineFrozen.step(16). No command66, New Game or Medium input followed. This failure was neither GC execution nor a native guest crash; the native admission fence stopped the command.

What ordinary pre-GC reachability actually establishes

OLD genuinely reached47d9c8 without injected New Game input. Its bounded stack positions6/14/22/39/42/45/47 contain 47ca7e/47a928/47b393/47c012/47c626/451546/402176. The original main-loop bytes show CALL451540 at402171 returning402176; main then increments/tests the periodic counter, conditionally PUSHes at402186 and CALLs443070 at402187, returning40218c. This saved maintenance chain supports admission before that invocation's subsequent conditional collector call. It does not prove the counter value, how soon that call will happen, or that no earlier collection has occurred during startup.

OLD command65 actually advanced16 steps to3600. Its ordinary New Game Enter down68/step69(16)/up71/step72(32) established full rest at3648, followed by a freshly reviewed difficulty dialog. Its first observer start was command77, after Medium keydown76. At that point actual PC449096 already had saved 44308f and40218c in the bounded stack. The observer started with zero observed boundaries and a sample; the subsequent gc return history missing fault at 40218c establishes missing entry history, not a newly observed collector entry. Medium's queued press was acknowledged, but native consumption and full rest were not established. This is why continuous observation was moved earlier.

The intended route is therefore: authenticate the original owning Module before instantiation; keep instrumentation dormant through ordinary startup; at an actually qualified native pre-GC checkpoint, start once before the existing observe16; checkpoint/resume the same observer across New Game press/release/full rest and fresh Medium press/release/full rest. Real entry443070 with [ESP]=40218c must precede its matching return at entryESP+4. Saved stack words never create those events. The published implementation tests this route with explicit synthetic producer/native fixtures. OLD proves the unobserved ordinary prefix reached its particular seam; neither original proves the continuous route completed or had sufficient50M/360s/512-record headroom.

NEW's4641c7 has none of the required seven stack-position values. Narrow original disassembly shows push 5 there, followed by an indirect call through 5c5794. That instruction is not the decoder output checkpoint. Static bytes describe possible next instructions; they do not prove that NEW executed them, returned to maintenance, or would next reach an admissible checkpoint. The 192-byte stack has no established maintenance chain equivalent to OLD's.

Cause boundary and smallest justified recommendation

The pinned runtime is Git1329b1ba6eb00144ff77d4cd92a8513d4fc953ad, not today's host.js. Its host adapts native work per slice using round((ranBlocks * 12 / ranMs) / 1000) * 1000, clamped between1000 and the configured ceiling, after sufficiently spent non-input-burst slices. Host steps therefore count scheduling opportunities, not a fixed native instruction prefix. Wall-time costs, yields and other Worker progress can change which guest instruction a given step reaches. The dormant wrapper forwards the original import and records run metadata; it has not activated the observer's per-boundary checks before this failure. Blaming active observer overhead for NEW's startup divergence would contradict that terminal report.

This mechanism explains why a step-indexed seam lacks a reproducibility guarantee. It does not identify the actual paired budget/timing sequence that produced these two PCs. Retained evidence lacks that sequence, cumulative native boundary history before admission, and the complete live stack/heap/ decoder state needed to replay either startup. The last NEW6000-boundary slice cannot fill those gaps. Matching controls, modules, screenshots or selected counters cannot supply the missing execution witness.

Keep the current rejection. The smallest justified next implementation work is bounded read-only provenance for ordinary startup slices and candidate native maintenance checkpoints, retaining budget, blocks, timing/yield, owner/epoch, registers and sufficient caller context. Qualify an actually reached seam before proposing any admission change. Any new route must prove that relevant tracked calls are not already pending, then witness GC entry/return continuously with one fixed contract. Do not add4641c7 to an allowlist, discard stack guards, silently run past the first failure to hunt for47d9c8, seed frames from saved returns, restart observation, reset deadlines or enlarge caps. Current evidence does not justify a production acceptance patch or another original attempt.

Validation: retained-data/source audit PASS; all seven relevant source/test files match exact repair71985c4506f6dd6caa374357e5812350cf1ced1d bytes. Reused logs are SHA/size-verified:178 integration tests,45 runtime tests, 17 focused tests and full locked build. No executable suite was rerun; docs-only git diff --check is sufficient here. No production code, runtime semantics, modules, shared files or queue locks were changed; no browser, game/module execution, heavy test or build was used. Menu captures and synthetic fixtures are not gameplay proof. Box ID/state/expiry:NA; no box was assigned or changed; coordinator owns shutdown.