# Jimmy special contacts: movement-stability audit 2026-11-04. Static recovery plus27 isolated original-dispatch observations; no runtime solver change. Current playable fccc0b9 remains the baseline. The user returned priority to Jimmy movement and explicitly allowed JN-Engine as a guide. JN-Engine's prototype movement is not imported or treated as an oracle. ## Result and immediate implementation boundary The lab/outside capsule bypasses material callbacks entirely: SceneCollision's private triangles retain geometric positions/normals but no linked material ID, and LabSimulation.StepPhysics receives only grounded/wall/ceiling flags. This is a concrete missing interaction layer, not evidence that changing walk constants or adding universal wall reflection will fix collision fidelity. **Do not globally admit lab/outside to OriginalBodyPhysics yet.** Besides dynamic pairs, three directly evidenced contracts are missing: material-driven hazard/ climb actions, world-response1D0 state during special animation, and contact-owned root publication. The existing ordinary sphere composition assumes no root-changing callbacks. Its wallContact hook alone does not establish the source's fresh previous/live Vec4 equality or special-animation lifecycle. For example, it currently adds the old falling displacement.Y after the hook returns a root, whereas the original can skip that entire candidate publication when the contact changed live/previous root. A root-changing callback cannot be made faithful simply by returning a different OrdinaryBodyState.Position. The smallest safe movement implementation chunk is ordinary ground/air arithmetic and phase stability while preserving the current route solver. For collision: first preserve/report actual contacted shape/polygon/3DMa identity in a diagnostic path; next implement the state-owning callbacks and source body modes with regressions; only then enable a bounded real lab/outside source-body route. Do not silently classify a special contact as ordinary to make scene admission succeed. ## Original identity and source map Neutron.exe SHA256: f5281076e1a2492f2d61f87aae3f4544b6d0f535ebe33f9fbb6f78710fe1aea8. Every code VA below maps to raw offset VA−400000, independently checked by PE. `tools/binary/jimmy_special_contact.py` records five bounded spans with hashes, raw offsets and four live virtual bindings. Existing material loader/source coverage is in collision-contact-callbacks.md and its guarded helper. | Code VA / raw offset | Observed contract | |---|---| |43AA40 /3AA40|Player contact handler takes **game interface** (Jimmy complete+C0), supplied Vec4 normal and polygon. First requires elapsed game774>3; missing polygon/material/chunk yields no special policy. Reads signed archive3DMa ID through polygon+4→material+C→chunk+8.| |43AA7E..43AAD5 /3AA7E..3AAD5|IDs120..129 call **complete virtual160**, IDs110..119 call **complete164**. These dispatches precede the movement-mode check and therefore still occur in nonzero modes.160/164 are slot offsets, not material IDs.| |43AAD5..43AAF9 /3AAD5..3AAF9|Fence IDs60..69 and ladder180 require mode game704==0. Other metadata returns without generic reflection/projection. The handler contains no raw controlled-pointer gate.| |43AB17..43ABBF /3AB17..3ABBF|Ladder first requires game338→472570 retained support polygon pointer (live234), then a47C210 probe. This is not the same predicate as current grounded218. A missing retained polygon prevents even the ray attempt.| |43ACD4..43AE19 /3ACD4..3AE19|Fence requires first probe to miss and subsequent side probes to hit. Miss-only synthetic probes make exactly two calls then reject. Different ray outcome sequences are not equivalent to a single wall hit.| |43AC29..43ACC5 /3AC29..3ACC5;43AF54..43AFDB /3AF54..3AFDB|Successful branches derive yaw from negative supplied normal through original angles, set destination/live orientation, transform local offsets, publish root314 and request LADDER/FENCE through complete150. Full successful-ray traces remain unexecuted.| |42A7E0 /2A7E0|Jimmy complete160 requires splat cooldown1E60<=0 and !1DED latch. Requests SPLAT via150, sets latch, clears impact clock1DE8 and controller scalar complete6C8/game608, stores supplied normal at1DF0..1DFC, requests−5 energy/score through42A920/42ADC0, and SFX37. Those helpers retain their own tutorial/global policies.| |42A8D0 /2A8D0|Jimmy complete164 requires !1DEC latch and raw gadget4F0588!=0 (Bubble0 immunity). Sets latch, clears1DE8, requests−5 energy/score, calls complete1E0→425120 (mode1 jump/effect cancellation). It does not share160's cooldown gate.| |437953..43796A /37953..3796A|Increment game774 elapsed **before** base40E050 and later raw ownership gate. Physics contact reads the preceding phase's elapsed. Omitting this timer would permanently suppress source policies; gating it only on Jimmy input is also wrong.| |43B408..43B4B4 /3B408..3B4B4;43B4BE..43B58A /3B4BE..3B58A|FENCE/LADDER animation entry owns saved destination Vec4 complete838..844/game778..784. Respective local offsets(0,320,150)/(0,800,150), velocity0, gameA8(false), action helper148, global busy4F83E0=true, duration complete888=4/6, fallback timer868=5, elapsed834=0, nonloop animation and SFX146/143.| |43A900 /3A900|AnimEnded handler returns FENCE/LADDER to STOP, clears busy/duration, calls Jimmy game410→42AFC0, clears action flag complete6E4/game624 and controller speed complete6C8/game608, restores saved full Vec4 root314, then gameA8(true). SPLAT/HIT branches differ.| |437988..437A15 /37988..37A15|Fallback timer game7A8 is decremented before input ownership gate. Expiry while busy can restore the saved root/response state too. Source duration4/6 and fallback5 are distinct; do not invent one universal fixed climb timer.| |47A123..47A1C4 /7A123..7A1C4|Body publication rereads previous2F8/live310 and compares all four components. A contact-owned root change skips candidate clamp/publication. Vector3-only stale equality is not this boundary.| Bindings independently checked: Jimmy complete4A3BDC+160→42A7E0,+164→42A8D0; Jimmy game4A373C+374→42A660→Player43AA40; Player game4AF6C4+374→43AA40. Game338→472570 returns retained polygon234, not a boolean grounded query. GameA8→470FD0 changes **world-response** getter1D0/live13C. It does not disable integration210, support1C8, raw control118 or global geometry4F69E5. Consequently `geometryEnabled=false` in the current managed body is not an implementation of this transition: that parameter follows global geometry-off gravity cancellation. The pose setter's exact callback acceptance, source morph/event cadence, impact latch/impulse updater4255C0, successful ladder/fence ray offsets/filters and reset interactions need component/runtime closure before actual mechanics are enabled. The stored impact normal is consumed later; immediate arbitrary velocity reflection would replace that state machine with a different behavior. ## REA and JN-Engine evidence JN-Engine docs/decomp/C3DPlayer.md, C3DJimmy.md and original decompiler bodies in docs/decomp/evidence/c3dplayer_movement_target3.md/c3djimmy_target6.md supplied function/field leads. Shared EXE identity is verified; its prototype movement and its different declared OMT2.dll build are not numeric/runtime authorities here. Class reports mix game and complete bases in places; the offsets above follow authoritative instructions and the independently established+C0 interface. Configured REA/Ghidra12.1.4 batch at artifacts/jimmy-special-contact/rea used the existing generated target copy and exported Evidence/closed snapshot: - 42A7E0: ev_801bbf1e21eb341f126f5e57e3efbfca42bb1c799680b2429c7f05dfdda3d625. - 42A8D0: ev_889e1c63d50f8e6da7e451ba0b821396f884fb7d7961dcdede3c4f3ca0e54af3. - 43A900: ev_0c030c12f6beb0790a6be2c53d3c4ac6591c8c2bd6552e8ea888f8de13455c7f. 43AFF0 returned an explicit procedure-recognition error; retained reply03.json is not successful decompilation and the unchanged selector was not repeated. Its supporting instruction extract is saved as special-animation.asm; the JN body is only a lead. Reused43AA40 Evidence: ev_4f2a6aa666451db831efd46f5c587573e22d8072315626b8bc2f844f9c6f0cae under artifacts/collision-integration/contacts/rea-callbacks. REA x87/return-address aliases are imperfect; reviewed call signatures/bytes settle the normal payload. ## Isolated original execution and tests The new generated probe executes copied43AA40 with synthetic game/complete vtables, linked polygon/material/chunk records, no-op160/164 callbacks and a miss-only ray stub. It **does not** execute real damage, successful climb, animation, body integration, collision traversal or the original game entrypoint. An initialized local Wine32 prefix runs the copied code in the project's separate network/filesystem sandbox; Wine exit0 and JIMMY_SPECIAL_CONTACT_COMPLETE are retained in artifacts/jimmy-special-contact/probe/{wine.log,capture.json}. 27 observations cover signed metadata and all dispatch boundaries, strict elapsed 3/above3, nonzero movement modes, null polygon/material/chunk, missing retained support polygon, and both hazard dispatch/fence ray attempts with raw controlNULL. Normal Vec4 reaches160 unchanged. Miss-only fence performs two probes; ladder one when support pointer exists. These are branch/gate observations, not terrain success or original end-to-end movement traces. `tests/test_jimmy_special_contact.py` commits the observations, original source span/binding identity and generator/harness/executable/log hashes. Four Python groups pass: source/offset/bindings, all27 native dispatch cases, provenance, and output/original/symlink isolation. Originals remain untouched. No runtime source or existing tests/docs were modified by this investigator; no commit was made. ## Proposed continuation gates 1. Keep source-backed controller/air/jump fixes and current playable route solver independent; preserve fccc0b9 and subsequent tested movement checkpoints. 2. Expose contact metadata/root-ownership diagnostics; recover4255C0 impact latch and impulse completion, then implement material120..129's bounded hazard flow. 3. Recover/verify successful climb rays, full animation-end/fallback ownership and independent world-response mode before adopting fence/ladder callbacks. 4. Complete dynamic pair eligibility/shared context and Vec4 publication, then run input-only lab/tutorial/RETURN/outside cases with known special contacts. General floor/step/wall fidelity is not established by these27 leaf probes.