NFS3 control consumption: static queue boundary; original attempt blocked
Run: scratch/runs/20261108-nfs3-control-consumption/ (ignored evidence).
Task incomplete. This run does not establish the current configured
binding, active player identity, input-caused acceleration/steering, or retained
road displacement. The sole fresh original attempt failed in the diagnostic
harness before catalogue completion, with sticky transport UNKNOWN. No emulator
mismatch or emulator fix follows from it. The earlier cockpit race and delivered
scan 0x48 evidence remain as described in
the previous investigation.
Original control path identified statically
Read-only objdump -d -Mintel of the original game/nfs3demo.exe is retained
as original-disassembly.txt. EXE SHA-256:
0defab3eeb22ee4b6e0007a4d5b26a99d868008ba77e2b9bd3ef770e924548ad.
These addresses describe the original executable, not emulator-private memory.
They identify where observations are needed; they are not execution traces.
| Boundary | Original code and data |
|---|---|
| Native default bindings | 0x48f720 copies defaults from 0x48ef10; the first four words are 0x4d04, 0x4b04, 0x4804, 0x5004. These are defaults, not measured current settings. |
| Configured table selection | 0x48efb0 selects 0x6fb240 when mode 0x6fb3b0 == 1, otherwise 0x6fb20c. 0x48dcd0 installs bindings through 0x48e420; 0x6fb4c4 can swap the first two assignments. |
| Binding to key array | 0x4314e0 extracts (binding >> 8) & 255, reads 0x562f28 + scan, and returns 0 or 255. The type dispatch table starts at 0x553ab4. 0x431510 also handles encoded keyboard bindings and additional fallback branches. |
| Poll to control packet | 0x48e030 reads installed bindings at 0x79a340, produces a signed axis byte at 0x79a338 and the next two channels at 0x79a339/33a, then packs button bits. 0x79a4ac is incremented by the poll routine; it is not a simulation-frame counter. |
| Packet admission | 0x48e4f0 has gates at 0x6fb3a0, 0x7d16a0, 0x7d16a8, and 0x55e818. Depending on mode it calls 0x4c4a30 with a player/input index, or 0x4c45c0. |
| Queue | 0x4c4a30 copies four bytes into 0x7d1714 + index*512 + (sequence & 127)*4; availability flags are based at 0x7d2714, sequences at 0x7d3718. It has explicit early-return/queue-full branches. 0x4c4d70 reads this queue. |
| Simulation-side snapshot | 0x4c4d40 copies four bytes from 0x7d16f0 + index*4; 0x48e180 uses it to populate 0x79a4a8..ab. This is a separate boundary from merely seeing a key held. |
| Car and playback selection | 0x42c110 installs callback 0x431140 at car offset 0x600 and 0x4bb880 at 0x5fc, with alternate callbacks selected by car flags at 0x200 and other mode tests. 0x4bb880 uses car input index 0x1f4; selector 0x7cba58 chooses 0x4bb500 for values 0/1 and 0x4bb590 otherwise. The former calls 0x48e180. The semantic names of other selector values remain unverified. |
| Car control fields | 0x431140 calls the car's 0x5fc callback, then copies 0x79a4a9/aa to car offsets 0x504/505, plus other control state. Car pointers are read from 0x5edac8; 0x5eda48 is also used by the selected-player path. |
The shipped readme lines 140–146 documents arrow/numpad equivalence; lines 289–290 uses Up acceleration and Left turning as its example. Static defaults agree with those scans. Neither establishes the actual configured table or which callback controls the visible car in a particular run.
The next useful observation is the first differing boundary along this chain:
current binding and player/callback identity, held key at an actual binding
read, generated packet, queue acceptance, simulation snapshot, and selected
car control fields. Capture the input index, queue sequences/flags and
0x7cba58 alongside car flags/callbacks; the prepared observer does not yet
cover every newly identified field. A moving opponent, attract/playback image,
or unrelated packet is insufficient. No focus, endpoint or missing-EXE theory
is revived here.
Local qualification and the concrete blocker
The released bound/observer helpers and complete registration closure were
reused as run-owned inputs. Guest and renderer module pin remains
7ec49cbbae4cb177909fd8fd9d3599713ee5f2a5ffb7f7e8c890848a540052bf.
local-diagnostic-qualification.json records generated x86 returning 0 with
the scan released and 255 held. Diagnostics on/off matched all nine recorded
registers and 128 output words; the original trace import ran exactly once per
observed event. This is a bounded synthetic semantic check, not proof of zero
runtime overhead or whole-game passivity.
local-generated-controls.json records actual Chrome trusted Up+Left events
through the exact eight-second bound helper, with independent reverse-order
key releases. A deliberately slow sampler separately passed release timing
and outside-hold labelling. Chrome exited 0 and its profile was removed.
Numeric/string endpoint rejection controls and all three registered-EXE
HTTP/SHA checks passed. There was no complete local staged-browser rehearsal
of the newly added observer attachment; that coverage gap mattered.
The worker's prepare.py incorrectly added control-observer.js to the
manifest with an absolute path. Upload and the browser's concatenated pin
check agreed on that wrong nested location, while the attachment code loaded
the intended run-relative path. The first attached guest Worker consequently
logged ENOENT at 06:39:04.348 UTC and remained paused before
Runtime.runIfWaitingForDebugger. At 06:39:24.028 the pending startup
Runtime.evaluate timed out with sticky UNKNOWN. Chrome closed automatically
and exited 0. This was my harness error; no configured binding, race-player
snapshot or original consumption trace was obtained. No input command was
issued, and the UNKNOWN was never replayed.
The failed runtime-pins.json is retained unchanged. The local preparation
path is corrected; staging-check.cjs and test-staging.cjs reject that exact
failed manifest and traversal/noncanonical paths and verify every byte in
runtime-pins-corrected.json. These checks passed after the failed run. The
corrected manifest was not launched. A future attempt must run this path
validation and rehearse the actual attachment path locally before allocation.
Evidence and normal shutdown
Shared production edits encountered during transfer were pinned back to the
recorded published blobs, rather than copying an inconsistent closure. The
runtime source commit is 0a38cf48438096482d059c5d8f27c048474c83b9;
base-commit.txt, relevant-source.patch, runtime pins and new-file patches
retain recovery inputs. Only small run-specific inputs were staged; media
was referenced by existing path/SHA. No source/media archive was made.
Both independent normal-stop and watchdog children were observably armed
before UUID-idempotent creation of bx_8ydmg3km at 06:38 UTC, with an immutable
25-minute deadline of 07:03 UTC. One heavy browser runtime was launched.
All 27 runtime artifact references (26 unique SHA values) were retrieved and
independently size/SHA-verified before normal stop.
The first drain assertion failed while resolving the same absolute pin. A separate settled assertion checked the actual staged locations and passed: 727 unchanged pins, controller and Chrome PIDs absent, both ports closed, profile and writer absent, and no matching processes. Its complete stdout was retained before stop. A second inventory request was refused locally because its attempt filename already existed; the standalone remote settled drain JSON was not fetched. No second remote inventory execution occurred.
Normal stop completed at 06:40:27.252 UTC, receipt
stop_0cfacaca207e1a87, forced: false, error null. Final box state is
archived, expiry null. Supervisor and both children terminated cleanly.
Disk stayed above 5 GB. No second box, resume/delete, guest patch, private game
call, clock/scene/config forcing, emulator source change or deployment occurred.
Full27, parity, audio and unmeasured FPS remain open.