Files
Neutron/docs/reviews/jn-engine-capture-request.md

74 lines
4.1 KiB
Markdown

# Proposed evidence request for jn-engine author
2026-11-03. Prepared locally from the supplied survey; **not sent**. This is an
investigation contract, not an implemented recorder/replay feature or schedule.
Our [assessment](jn-engine-survey.md) explains the evidence and version limits.
## Existing material first
- Analysed Neutron.exe Ghidra project/export, Ghidra version, analysis options
and applied script list. Saved target6 bodies are already useful; preserve
prototype warnings and explicit function boundaries in the handoff.
- Raw Level1 OMTC for anchor8881 and constituent world-fixture keyframes, with
matching original screenshots and fixture-generation commands/settings.
Derived scenes/draws/textures exist locally; raw capture/reference image do not.
- Hashes of the actual captured EXE, OMT2.dll and proxy, executable variant,
display/renderer settings, starting level/save/task state and capture build.
EXE manifest matches ours; OMT2.dll manifest currently differs.
- The offered written code/script permission or permissive licence, if direct
source import is intended. Keep original game assets/binaries distinct.
## Minimal original-state recording
First record a short VR04 standing/walk/release/turn/jump/held-jetpack sequence.
Then lab→VR04→lab→outside, including intro, Carl control handoff and RETURN.
Gadget activation, death/restart and save/load can follow a validated basic trace.
Please record a versioned little-endian schema and an address/base map. Every
state field should name its width/type, pointer base, sampling phase and source
instruction/helper; mark unresolved fields rather than guessing. Include:
- Update sequence, render sequence, phase, game dt and wall timestamp separately.
Renderer EndScene timestamps alone cannot establish simulation cadence/order.
Identify skipped/invalid-dt and paused updates; report dropped frames/records.
- Initial save/TSK/checkpoint/level/start identity, RNG and relevant settings;
input down/edge state, focus, mouse/controller axes and polling phase.
- Raw controlled pointer and class identity, independently tracked Jimmy identity,
pose/velocity, movement mode, jump phase, support/shadow distance, pending
impulse, control/cutscene owner, gadget display/raw mode and energy/timers.
- Camera mode/owner/transform/clips, task/SCENE changes, objective/UI transitions,
audio start/stop identifiers and level-load/RETURN/reset boundaries as feasible.
Start with samples immediately around the game-update body, then add subphases
where a discrepancy demands it. Our current static map has dt write47E502 to
696030, physics47E554, registry47E582→473740, transform sync47E587→473770,
base draw47E6DF and render counter increment47E6EF. These are **EXE virtual
addresses**, with their current evidence in `../checkpoint/loop-motion.md`;
validate safe hook boundaries/ABI before instrumenting a copied installation.
Registry order and changes within the update can matter, so a lone end-of-frame
snapshot cannot establish which owner wrote a field first.
## Pointer-base correction needed before capture
Jimmy's game interface is complete object+C0. These established pairs are:
| Value | Complete Jimmy | Jimmy game interface |
|---|---|---|
| Movement mode | +7C4 | +704 |
| Jump state | +7E8 | +728 |
| Shadow ground distance | +86C | +7AC |
The survey's proposed list mixes the two columns. The raw controlled global
5099E4 can be null or owned by another class during cutscenes/vehicles; do not
apply Jimmy offsets without class validation. Root/velocity access slots are
virtual getters, not established raw memory field offsets. Resolve actual
storage or use a reviewed getter strategy without changing simulation state.
Similarly verify widths/layouts for proposed SCENE/objective/gadget/RNG taps;
the survey alone does not define a serialization contract.
Acceptance: one short recording with reproducible startup, explicit schema,
complete input/dt/state phase alignment and no unexplained record drops. Compare
that recording manually before expanding recorder coverage or implementing a
GamePort replay consumer. Keep capture/replay work local with separate generated
outputs and untouched originals.