Alice original host receipt, ground capture and route — 2026-11-03
Run: scratch/runs/20261103-alice-host-receipt-ground-route/ (ignored evidence).
This is one task on one fresh boat, bx_974w85x2, originally expiring
2026-11-03T20:09:42.170Z. No predecessor boat was operated.
Exact input and evidence boundaries
The staged tracked input is commit 277f9a32c984562ef33b2d32a5c13df949b5f87a,
including Alice fault retention 0c2bc2da5641695157703a778a377e6f44b16d74.
The matching-input check refused reuse of the older 0a58fd7b… module:
compile inputs differed. Compilation of the exact staged input produced
2,125,777 bytes, SHA256
065267ce7c208cb369b1fa930425627b3a238c9e184e5aa46d8ef16143b8b5b8.
The 43 original/dependency pins, actual game manifest, pngjs, stdole2.tlb,
imports and startup were verified before original runtime. No compilation
archive was downloaded. Later sessions reused this module only after unchanged
compile-input and full source/module verification. Each process has its own
host module receipt matching compiled and instantiated bytes and PID.
Inherited post-stop capture/contact/request/failure repairs, startup readiness,
and module receipts were pinned from the three predecessor repair manifests.
The old accepted cave/menu/readiness fixtures were not repeated as acceptance
work. Every new session used original NewGame/Easy and the natural opening.
The control entrypoint waits for the same live PID/start tick and the actual
server.listen callback before command one. Module bytes are read from host
receipts; no actual_module/wasmBytes evaluator submission was made.
First failures and bounded repairs
Session one (PID 6511) reached the original cave at batch 4104. Its first
movement-return capture failed movement argument record mismatch. Full raw
native module/lifecycle tables, stack, selected Player diagnostics and first
error were retained. The reserved world classifier had not yet executed.
Original callback instruction 10097a7e: mov [ebp+8],ecx overwrites the first
incoming argument with an event temporary. At cdecl return 00452f63, that
stack slot therefore cannot be required to equal the incoming record. The
second stack slot remains the command pointer. The first attempt did not read
those pointed command bytes before failure; they cannot be reconstructed from
the separately retained engine copy.
The task-local repair retains the complete return stack and event scratch, reads the actual pointed 20-byte command and independent engine client copy, and requires byte equality. It does not invent incoming argument bytes. Synthetic controls, the actual retained first stack, and the original mapped overwrite instruction test this repair; these are offline controls, not new native qualification.
Session two (PID 13919) reached the natural cave but an optional pre-call
breakpoint targeted instruction 00452f60 inside a decoded block. One credit
completed with the gate still armed and unheld. A host assertion expecting a
hold failed; no native fault was reported. The original gate could not replace
an unheld arm. The session was stopped ordinarily with its diagnostics intact.
The second task-local repair permits replacing an unheld arm after validating the replacement signature. Held/failed gates remain immutable. Five focused controls passed locally and on the same boat, including invalid replacement preserving the prior arm, held replacement rejection, normal release and duplicate release. Session three (PID 21137) uses this tested gate and the command-stack repair. No tracked runtime, compiler or Jedi source was changed.
Native outcome
Session three is PID 21137, session
b29fff6e-9f1e-4876-9c59-8d00131eda90, actual DLL base 0x02d40000.
Native held movement-return captures passed 56 relocated windows and 377
relocation operands each. The complete loader table (capacity 64), lifecycle
table (64-byte stride), raw stack and command copies accompany each capture.
| Held epoch / batch | Origin | Ground number / walking / plane | Contact |
|---|---|---|---|
| 1 / 4106 | (-4183.4814, 5999.9497, 32.125) | 1022 / true / true | reserved world, water 0 |
| 3 / 4147 | (-4178.5610, 5993.1230, 32.125) | 1022 / true / true | reserved world, water 0 |
| 4 / 4176 | (-4178.2803, 5972.3833, 128.2834) | 1023 / false / false | null ground/contact, water 0 |
Epochs 1 and 3 have identical neutral command input fields, with different command times; there is small XY drift, so exact stationary coordinates are not claimed. Epoch 4 follows ordinary Space input and retains its actual 20-byte command including upmove 127. Airborne classification follows native flags/records and height, independently of screenshots. The actual API has live count 208 and capacity 1024: reserved world 1022 is handled separately from ordinary live entity bounds, repairing the original rejection against live 208.
Epoch 2 stopped at the actual DLL callback entry. Its raw stack contains
return 00452f63, selected record 7ddb0014, and the actual command pointer;
the pointed command and engine copy agree. This entry and epoch 3 are different
movement invocations. They are not presented as a paired call/return trace.
All three return captures select Player 03aea628, raw allocation 03aea614,
in the same PID/session/module generation and backlink chain. The passive
native allocator log has 106,170 events through the first route movement,
zero observer errors, and one matching allocation: sequence 88,497, batch
3302, requested 2048 bytes/tag 4, aligned block 2068 bytes. No matching free
occurs through the sample interval. Three same-held header/tail reads match
the block size, DLL size, sentinel 000facc2, and relocated Player vtable.
Native allocator/free, DLL wrapper, deleting destructor and vtable bytes match
the exact original PE bytes with relocations applied. This establishes the
same observed Player allocation lifetime across the samples. The original
capture helper's conservative allocationGenerationProven:false and
usableForFooting:false fields are preserved; the separate
attempt3/native-sample-assessment.json records the joined native evidence.
Ordinary W, strafing and relative mouse input traversed the cave/timber bends
and reached the first gnome's “Ride the wind” sequence, observed by batch 6238.
The inspected capture at 7838 shows the HUD returned. Native epoch 11 then
records (-4974.8232, 5419.0796, -119.875), heading -180, dry reserved-world
support. The subsequent west/south approach overshot into the steam basin:
epoch 12 is (-5165.0708, 5061.2246, -379.4069), water level 3. This is below
the authentic wind trigger's lower bound -320. One ordinary Space attempt
does not establish a lift: epoch 13 remains underwater on the basin floor,
(-5172.9214, 5077.1572, -383.875). The timber plank and steam were inspected
in screenshots, but wind lift, crossing, far-side landing and knife acquisition
remain unverified. No attack was dispatched.
At closeout, an avoidable auxiliary input-state request was rejected because
the launch did not enable --trace-input-dispatch. Its first error is retained,
and it triggered ordinary exit 0 after 1479.511 seconds. Final fault-status and
snapshot preceded that rejection. The native samples and gameplay above were
already complete. A separate offline closeout-plan guard removes that request
for this launch and tests missing/declared flag cases; it is not native-qualified.
Another full natural opening plus recovery could not fit the remaining lifetime; none
was launched just to repeat this auxiliary diagnostic. Read-only retirement
checks confirm game/supervisor/helper PIDs absent and control port 8123 closed.
The final native fault receipt is armed, failed false, terminated false, fault
null. It distinguishes the host diagnostic rejection from a native guest fault.
Recovery and shutdown
The run's RECOVERY.md specifies complete source/module/evidence restoration.
Large payloads remain in one complete archive rather than duplicate binary
trees. Every archived payload is streamed through size and full SHA256
verification locally before the explicitly delegated ordinary boat STOP.
Synthetic recovery controls reject changed bytes, duplicate members and path
traversal. The complete archive contains 7,456 payloads, 168,726,074 bytes,
SHA256 3ce6777c1245653c68b02e0535019b757b7c049356b65a5f1a4f4f6953b6af34.
All payload sizes and full hashes passed locally before STOP. Final source
checks passed 5,931 / 5,964 / 5,964 pins for sessions one / two / three.
Boat state is confirmed stopped. Stop stop_00dc1ca82c143018 completed
2026-11-03T20:00:35.847Z, error null, forced false; the available snapshot
completed and verified at 20:00:33.613Z. No resume or successor was created.
Only this note is committed; executable repairs and full evidence remain in the ignored run and complete archive with recovery instructions. No tracked source changed, so a new repository build was not required. Focused scratch controls, actual original compilation, native captures, full recovery checks and documentation whitespace checks are recorded separately. Remaining gameplay limitation is the basin-to-wind crossing, followed by the far-side plank/knife route; neither crossing nor weapon readiness is claimed.