# Carl skip lifecycle: recovered boundaries and the remaining observation 2026-11-06. Execution of the accepted next-work plan, with the user's requirement to preserve original behavior exactly. Runtime a2e89c6 and packages dae13ac remain unchanged. This checkpoint adds evidence and observation tools, not a tilt fix. ## Event/state trace Original source identity, PE offsets and REA Evidence are in the linked reports. These rows combine static facts and explicitly synthetic original-code execution; they are not a captured whole-original frame sequence. | Boundary | Recovered state ownership | Evidence / remaining limit | |---|---|---| |Dynamic contact→Carl state0|Matching target and Carl state6 admit the talk-ready transition. Jimmy is not modified and support is not queried by that callback.|[Contact report](jimmy-carl-entry.md):96 native cases; actual physical trajectory/admission and whole AI update are separate.| |Carl talk→SCENE50→60→neutron1a|Task assignment/broadcast precedes the synchronous camera trigger. Jimmy's authored TaskName is `none`. All89 authored SCENE listeners across17 kinds are inventoried; six callback paths do not supply another Jimmy movement reset.|[Task report](jimmy-task-return.md): static callback closure and48 native dispatcher/objective cases; actor callbacks in that probe are adapters.| |MCA initial preparation|StopMotion clears speed/velocity independently of STOP animation acceptance. Both current and destination Euler become(0,currentYaw,0); dirty is preserved. Raw control is removed.|Earlier20 camera-entry cases; [full STOP/base report](jimmy-base-return.md) supplies the omitted selection/completion dependencies.| |Uncontrolled interval|Known controller ownership gate retains mode1/stored turn. Actual base/facing and ordinary STOP/talk/FALL completion provide no reset. Playback/resource state is distinct from movement mode.|140 native base/STOP cases. Full body context, complete Jimmy frame and OMT playback remain outside these probes.| |Accepted skip|Original sets Jimmy action/menu latch complete1E00 and shared skip4F83E4; stops sound, deactivates current actor/animation and invokes completion.|[Return source](jimmy-cutscene-return-source.md); actual skip prefix is executed by the composition. The complete skip frame is not.| |MCA completion→Carlwalk→DEFAULT|Completion grants named JIM1; Carlwalk changes Carl's AI. DEFAULT's already-owned early branch clears shared skip/resources, with no new Jimmy reset.|[Complete DEFAULT recovery](jimmy-default-decompilation.md) closes the previous provider-boundary gap using an explicitly mapped static carrier.| |First resumed callback|On synthetic supported/released state, input-before-player republishes stale roll before the controller clears mode/turn. Player-before-input clears those fields first and stays upright.|[Composition](jimmy-skip-composition.md):8 actual original-component cases agree with48 port traces. This does not establish the original route's actual entry state/order/delta.| At the problematic composed boundary the seed is mode1, turn±11.666668, current/destination(0,155,0), supported=true and dt1/60. Fresh input reduces turn to±10.000001 and publishes the opposite roll. The next controller clears mode and turn; neutral ground input then retains the banked destination. Skip-prefix latch writes do not change this result. These values are measured synthetic component inputs/results, not values measured from the original Carl route. ## What this permits and what remains unresolved The first stale port writer is known. Its arithmetic also occurs in the composed original routines, so adding a reset to raw control grant, DEFAULT, Carl contact, task broadcast or ordinary animation completion would lack source support. The user's observation that the original stays upright remains authoritative; the actual original state/order entering that first input is still missing. Normal landing bank must remain intact. No per-frame leveling, scheduler switch or input suppression has been introduced to make the diagnostic green. The fresh-construction order is a conditional source fact, not a runtime trace of the user's route. Startup inspection constructs MAIN directly and supplies no positive lab-preload/reusable-Jimmy explanation. A new read-only observer can record real registry positions, raw ownership, mode, turn, current/destination Euler, dirty, cached support, viewport delta and input/skip flags. Its samples are non-atomic and cannot alone certify exact callback chronology. ## Runtime observation tool and local limit [Original-state observer](original-state-observer.md) supports launch or attach to a verified generated game copy. It never writes actor memory or injects code. The attach path leaves the observed game running. Three exact executable profiles have24 independent instruction guards; the supplied HDnoCD variant has its own REA Evidence and global addresses, rather than assumed canonical equivalence. The local Wine32 attach integration passes:928 parsed samples, source guards accepted, observer exit0, original still running after sampling. Provenance is `artifacts/original-state-observer/attach-runtime/observations.json`. The legacy window remains black; observed registry contains MAIN and no Jimmy, with viewport delta0. Focus/normal Return input and an isolated desktop also failed to reach a usable gameplay reference. Neither accepted input events nor MAIN's zero gameplay delta proves a menu selection or a whole-game timing fault. Native Windows and the actual Carl skip were not observed by this work. To collect the missing comparison on native Windows, run a **copied** original, reach Carl, then attach for the banked entry/skip/neutral-return interval: ```text observer.exe --attach-copied-pid=1234 --seconds=60 > carl-original.jsonl ``` Use the copied game's actual PID. Keep its path outside an `Original` directory; the tool checks the actual executable identity/path and refuses unsupported images. No automatic keys or close request are sent in attach mode. If the trace misses the first transition, a narrower callback-level observation is required; do not fill it with an inferred reset. ## Checkpoint status 292 new bounded original-code cases pass:96 contact,8 input/controller/skip-prefix composition,48 task/objective and140 base/STOP. They deliberately have different synthetic dependencies. The48 port diagnostic cases still describe the preserved failure, not successful gameplay. The mapped DEFAULT carrier regenerates exactly; 539 original files remain unchanged. No src/tests/scripts or package change is part of this checkpoint. Steps1–2 of the next-work plan remain open at the actual original resumed-frame observation; step3's gameplay correction gate is not met.