Files
Neutron/docs/gameplay/jimmy-rotation-handoff.md
T

6.1 KiB

Jimmy rotation at script and ground handoffs

Follow-up: the entry fix below does not resolve neutral idle after skipping Carl. User reports that separate port bug and confirms original landing bank is normal. See decompilation checkpoint for the 48-case reproduction, original skip findings and the implementation pause.

2026-11-05. User reports jumping at an angle into cutscene triggers leaves Jimmy tilted, and a slight tilt can also remain after landing and standing still.

Recovered cutscene correction

Original camera ownership records REA Evidence, original addresses/file offsets, callers, vectors and native execution. CAM416671..4166EB and initial MCA431518..431590 explicitly flatten both current Euler and facing destination, preserving current yaw and the dirty flag. Their direct game+274 writes were previously described as a render cache; the air bridge correctly identified them as destination state but missed this handoff. The port only flattened current Euler. Control-independent facing then consumed the stale airborne target and restored the tilt (and sometimes the prior yaw).

LabSimulation.FlattenOrientation now replaces currentXYZ and destinationXYZ with (0,currentYaw,0) without calling the dirty-setting destination setter. Existing CAM/MCA preparation ordering, prior-owner guards, AI NULL/DEFAULT differences, held input, queued motion and fire lifecycle remain intact. Authored transforms still change current Euler independently of destination; no generic control-loss or per-frame orientation clamp is introduced. Original orientation W is verified by the probe but remains outside the port's XYZ orientation model.

Landing investigation

Landing source investigation records its separate boundary. Neutral ground input retains its prior destination; ground apply, STOP, the landing notification and the derived impact latch do not supply a proven leveling reset. This differs from the directly evidenced camera reset. Do not invent an unconditional grounded snap from this symptom. Seven copied-original300-tick ground-input/facing trajectories reproduce retention: destination (0,155,-5) retains roll355 after300 neutral passes; active Up/Left replaces X/Z targets and levels. The managed ground tests compare all seven source trajectories. These are component observations, not a whole original frame trace: another unobserved frame writer or an original quirk remains open. No landing policy is changed without that evidence.

Verification

The old implementation fails the new director regression at the pending target. Eight integrated cases cover CAM/original lab Carl MCA, both input/player orders, and clean/dirty facing. They enter real banked jumps, verify immediate current/ destination flattening and unchanged dirty/impulse, then run120 locked body ticks without reintroducing tilt. Additional checks compare20 saved native camera preparation/facing XYZ/dirty outputs, including negative/wrapped Euler and yaw seams. These comparisons seed dirty through live input; the native probe's whole object comparison separately proves independence from the old destination and W. The saved fixture includes original range and probe provenance hashes.

Native Linux SDL/Vulkan --smoke-control-phases now first creates a real8-tick banked jump before triggering the original Carl MCA. It asserts both current and destination are leveled while dirty survives the handoff, verifies they remain upright through15 locked body/fire/energy ticks, and checks held-input resumption. It saves locked-start/locked-fall GPU captures. The Carl lock and powered VR04 replays remain regression checks. Artifacts: artifacts/jimmy-rotation-handoff/. 28 movement groups, extended Director tests, original25 ground observations plus seven trajectories,14 jetpack groups and both visual groups pass. Build passes with zero warnings/errors using NuGetAudit=false (the first default-audit attempt reported unavailable network metadata). Native Linux banked handoff/Carl lock/ powered jetpack runs pass with zero Vulkan validation errors. The jetpack replay requires explicit VR04; an initial default-lab invocation was rejected by its existing argument guard before rendering and the corrected replay passes. All336 Python tests pass, including new camera/landing source and native-fixture provenance guards. The initial managed fixture name collision was corrected before the final Director run. Package results follow below.

These are static-source, isolated original-component and port-runtime checks. They do not establish whole-original frame scheduling or native Windows parity.

Packages and final boundary

Both self-contained packages identify clean runtime commit dae13ace7391671aee7082c662bb798b285b9124. Each verifies519 original copies, 4285 prepared scene/resource files and seven prepared audio files. The539 source originals remain unchanged. Provenance is in artifacts/jimmy-rotation-handoff/package-identity.json. The earlier turn/fire package resource count likewise excluded those audio files; no new assets are introduced by the orientation correction.

Standalone Linux from /tmp with SDK variables unset passes the banked Carl handoff/body-fire continuation/control-resumption replay with zero Vulkan validation errors. All three packaged handoff GPU captures match development byte for byte. The three powered-flight development captures also match the previous turn/fire runtime exactly. Hashes/logs are retained as package-readbacks.json, package-control-phases.log and jetpack-readbacks.json. Windows is cross-published/resource-verified; native Windows and whole-original frame comparison remain outstanding. Validation docs are added after packaging; BUILD-INFO preserves the clean runtime commit above.

The landing report remains an investigation outcome, not an implemented landing correction: verified components preserve neutral bank and do not provide the missing whole-frame observation. Any automatic landing leveling requires further original evidence or an explicit decision to diverge from the recovered behavior.