Maelstrom startup divergence: trace fanout changed the execution budget
Historical source last updated, 2026-10-24. Runtime and implementation claims below describe that dated source; see current app notes and source for subsequent changes.
The earliest provable experimental divergence occurs before guest startup. Canonical runner test/run.js parses five --trace-at addresses and unconditionally assigns BATCH_SIZE = 1 at lines 1528–1530. The preceding reached-Sync run specified one trace address and retained the default 1,000 at line 288. The main loop calls instance.exports.run(BATCH_SIZE) at line 10363. Both runs kept 42,000 batches and identical batch-number input/clock schedules. The later capture therefore did not preserve the execution prerequisite under which the earlier capture reached startup, afterNextWave and Sync.
This is a concrete harness cause of unequal execution budgets, established by evaluating the exact hash-checked startup statements without loading WASM. It is sufficient to reject the latest recipe as a passive follow-up to the earlier run. It does not prove that budget restoration alone guarantees the same entire guest outcome, or identify the precise incoming native call at the final PC. No game, network, interpreter or _ftol defect is established.
The checks (historical artifact: maelstrom-startup-divergence-cause-20261024/checks.json) retain exact source slices and fragment hashes. --trace-at-start-batch=36163 delays breakpoint arming at lines 10247–10258, not the startup size assignment. --trace-from=30000 gates logging, not execution size. --trace-at-limit=16 limits records, not the earlier assignment. Even explicitly adding --batch-size=1000 still yields size one because trace fanout overrides it. The rejecting controls preserve this failed prerequisite.
42,000 calls with requested budgets of one versus 1,000 are upper budgets of 42,000 versus 42,000,000 blocks, not measured retirement totals or a claim that total work differs by exactly 1,000. Yields and cache work can shorten execution slices. The clock's tickState.batch = batch at line 7527 and the fixed 200ms-per-batch schedule continue unchanged; the P key still arrives at batch 31,000. Thus CPU work available before each scripted input and capture also changed. Silently extending the batch ceiling would change the experiment again and is not a correction proposed here.
Native boundary containing 0x417430
Authentic PE parsing independently derives image base 00400000, entry 00417474, section raw mappings and import lookup. The six original bytes at 00417430 are ff 25 d4 90 41 00, JMP DWORD PTR [004190d4]. That import slot names MSVCRT.dll!_ftol, the floating-point-to-integer conversion helper. It is not a startup/exit API. The nearby 00417474 is the separate CRT entry prologue; 00417529 calls __getmainargs, and 004175a3 calls 00417860 before the exit call. The startup window (historical artifact: maelstrom-startup-divergence-cause-20261024/startup.asm) and tail (historical artifact: maelstrom-startup-divergence-cause-20261024/startup-tail.asm) show the boundaries; byte checks reject a changed IAT operand and changed import name.
Eight original direct calls to that trampoline are corroborated in the three narrow caller windows. Their real return PCs are 40387c, 403886, 403890, 40389a, 415790, 4157f3, 41638b, 416414. The final latest log has ESP=074ffbcc, but does not retain [ESP]; consequently none of those callers is selected from the final registers. Final EAX=1 is not an observed _ftol, select or CheckSockets return. Final EIP at a conversion trampoline is consistent with a bounded stop during startup work; it does not by itself prove a stuck helper or an asset lookup failure. The exploratory caller-A slice started in a previous instruction and remains charged at recovery root; the committed replacement starts at aligned 403857. No conclusion uses its spurious leading decode. Every native window is below 8KiB; all task disassembly, including that retained exploratory slice and replacement, is below 64KiB.
Exact comparison and retained outcomes
The logs are byte-identical through working-directory setup and the trace-window marker. The first retained output divergence is the earlier run's SDL CreateThread start=688c40, absent from the latest run. This is later than the proven runner configuration divergence and does not locate the first different native instruction. The previous 6,947-byte output, command and terminal are byte-equal to their immutable 6c265033d01e547a3d9320316cc17027bbb7cb0b copies. It reached 406378 at batch36163 and Sync once, with 972 select returns at41999. The latest 3,897-byte output, 2,088-byte command and 309-byte terminal, plus all three PNGs, were freshly verified against all six accepted raw seals from donor 06d9827805e8eb191c4d7ce2ca43217db413f523. All six latest counters are zero, final EIP 417430, wndproc zero, 4,702 APIs and 42,000 batches/0.428s. The coordinator already inspected the three identical uniform-teal PNGs (SHA256 cf54108b7650531aeaa719b810d53cd2c624323b760ba730b7897261efad2d68); this task rehashed them without new image interpretation. Natural host return zero is not guest normal exit. Existing failed runtime predicates remain failed; no Sync packet inference is added.
Runnable checks and concrete prerequisite change
timeout 60 python3 tools/maelstrom-startup-divergence.py --output scratch/runs/<run-id>/maelstrom-startup-check.json
timeout 60 python3 tools/maelstrom-startup-divergence.py --check-argv scratch/runs/<run-id>/command.json
The unexecuted single-return argv (historical artifact: maelstrom-startup-divergence-cause-20261024/single-return-argv.json) changes only --trace-at to one real CALL-return PC, 0040e12a, restoring startup size1000 under the exact canonical parser. It retains all other latest flags, including 42,000 batches, the clock and input schedule. Its offline guard passes. This is a concrete distinguishing prerequisite for a future coordinator assignment, not authorization to run it. Do not add other trace/break/watch addresses that re-enable fanout. The existing real 406378 counter is the smallest startup reachability observation: positive after budget restoration reestablishes the prerequisite for packet diagnosis; still zero means startup remains unresolved and no Sync return claim is justified. At the single selected real return PC, observe EAX and existing ESP:384 memory (saved missing count [ESP+18]); absence of that trace is not a timeout result. No longer run or network change is proposed.