Trace post-skip tilt and gate gameplay changes on original lifecycle recovery

This commit is contained in:
UnixNight committed 2026-11-05 22:14:49 -05:00
1 parent a2e89c625a
commit 68154dffce
15 files changed
+752 -6

No files matched your search

@@ -0,0 +1,65 @@
# Laboratory trigger and skip callback order
2026-11-05. Static original source/data recovery; gameplay changes are paused.
Original EXE identity and retained REA Evidence are recorded in
[jimmy-cutscene-return-source.md](jimmy-cutscene-return-source.md).
`tools/binary/jimmy_cutscene_chain.py` verifies PE-mapped offsets, calls, virtual
bindings and supplied GAM fields; generated output is
artifacts/jimmy-cutscene-return/source-contracts/chain.json.
## Actual lab records
| Record / file offset | Source chain |
|---|---|
|3MCA neutron1a, GAM index7, E0B|Five active targets, beginning with JIM1, then Carl, Carl, Jimmy, Carl. PlayerControlledJIM1 is the completion target; DeactivateInv1; TargetDeactAnimSTOP; NextTriggercarlwalk.|
|3AIT Carlwalk, index10,1AFD|AITargetC3DCARL, AIState3, no new position/yaw, PlayerControllednone, NextTriggerdefault. This record does not restore an old Jimmy transform.|
|3AIT teleportexplanation,index34,6756|Touch activation byjim1, SCENE30 invokes te1; PlayerControllednone; te1→te2→TE3→te4→te5→default. This is a different entry path from Carl's friend talk.|
GAM hash316fff40db565c2add5b07cbccdef640ddb67c64d6625fea1429ce2a0542a234.
All tags are looked up case-insensitively as in the existing bounded parser.
## Original partial order
Viewport47E554 calls global physics479BC0 before registry473740 at47E582.
Physics invokes dynamic pair processing47A430 at479C83. Its enter callbacks
47A555/47A578 dispatch slot40 before the later registry traversal.
- An AI trigger's interface is complete+C8:48F8C0+40→40C200, then40C21F
synchronously dispatches40C300. A trigger can therefore activate a camera
during the contact pass, before the game's ordinary input callback.
- Carl's interface is complete+C0:495364+40→40A3C0. Its AI contact state may
become0. The friend update41B950 first calls408000, then41B972..41B988
invokes talk when the resulting state is0. In the fresh registry partial
order, Carl follows Jimmy. Contact and actual conversation activation are
separate boundaries.
- Talk41B7B0 obtains a matching trigger, invokes Carl's progress callback,
then41B7D8 dispatches the chosen trigger. Carl413E50 calls41B810 first
(+100 score and **Carl's**410 preparation), then advances SCENE50→60.
This does not by itself establish a Jimmy movement reset.
- MCA update is game interface+C8 table4A8F64+3C4→430DA0. Natural completion
and skip have different branches; skip writes Jimmy's action latch and
shared skip flag before completing through4A93C8+E4→4311F0.
- Source registry473740 dispatches each object's slot10 at473758. Actor
insertion order matters for same-frame control changes. Fresh and retained
routes must not be collapsed into a single assumed order.
The port's director.Touch/SceneInteractions calls occur after the combined
player update in Program. This is a documented scheduling approximation; it
cannot prove the original starts a conversation with identical support/mode/
turn state. The original dynamic admission/support and callback positions need
comparison before that difference can be called the tilt's source cause.
## Missing proof and next trace
The user specifically reports **skipping Carl**, not a measured natural-end
comparison. The diagnostic also reproduces natural-end failures in the port,
but that is not evidence about original natural completion. Required original
trace points are: physical contact, final Jimmy callback before activation,
MCA entry, accepted skip, completion/DEFAULT, next game input and next Jimmy
update. At each point record raw owner, support, mode, turn, current/target
Euler, dirty and input suppression/latches. This will distinguish an entry-state
mismatch from a return/scheduling mismatch.
Current static conclusions and managed traces are useful evidence, but neither
supplies a source-justified reset location. No gameplay change is introduced by
this report or its read-only guard.
@@ -0,0 +1,80 @@
# Focused decompilation checkpoint: cutscene lifecycle
2026-11-05. User redirects work toward understanding the original before further
implementation. Preserve runtime baseline a2e89c6 and packaged runtime dae13ac.
No src/, regression test, build script, original or packaged binary changes are
part of this checkpoint. New work is read-only diagnostics and source findings.
## Confirmed requirement and corrected coverage
User comparison confirms ordinary landing can leave Jimmy banked in the original.
The reported port-only failure is specifically after **skipping Carl's lab
conversation**: Jimmy must remain upright when control returns, including idle.
The earlier correction proved camera-entry current/destination flattening and
locked continuation. It did not prove resumed **neutral** input stays upright.
Its held movement/jump resumption check cannot certify that missing boundary.
The post-skip bug is open; do not classify it as the original landing quirk.
## Findings this pass
- [48-case port trace](jimmy-cutscene-port-trace.md): all24 fresh-order cases
republish a stale air-bank target on the first neutral input; all24 retained-
order cases remain upright. The player is grounded by return. This isolates
the first port writer, not a source-justified fix.
- [Original return/skip boundaries](jimmy-cutscene-return-source.md): raw control
grants, DEFAULT and no-selection cleanup do not supply the hypothesized mode/
turn reset. Original skip does write Jimmy's shared input latch and a global
skip flag, with per-camera completion/resource behavior absent from the port's
immediate fast-forward model. Its causal role in upright return is unproven.
- [Source chain and registration order](jimmy-cutscene-chain-order.md): the actual
GAM chain is neutron1a→Carlwalk→default. AI trigger contact can dispatch during
physics; Carl contact first sets state, with conversation activation in his
later actor update. The port's post-player contact grouping is an approximation.
- Three fresh REA/Ghidra pseudocode results and the existing full instruction
Evidence corroborate the focused source contracts. DEFAULT's unrecognized
procedure remains an explicit provider gap; assembly/byte guards support the
bounded statements. Complete Evidence is retained locally and sessions closed.
These findings improve the interaction model. They do not establish that the
original enters/ends the scene with the same support/mode/turn values as the port.
No additional rotation reset or blanket ordering switch is implemented.
## Recovery gate before implementation
1. Trace the original Carl route from physical enter to talk activation, then
accepted Escape/Space skip, completion/DEFAULT and the first idle callbacks.
Record support, raw owner, current/destination Euler, dirty, movement mode,
turn, shared skip, action latch and relevant input suppression at each boundary.
2. Close source actor/input ordering for the observed fresh and retained route,
including deferred callbacks and contact/support eligibility. Explain where
the source avoids the demonstrated first stale write. A local no-reset finding
or a synthetic component outcome does not substitute for that explanation.
3. Compare a bounded original composition/runtime trace with the reproducible
port failure. Keep original instructions, inference and user observation
separate. Cover skip and natural end independently; user observation here
specifically concerns skip.
4. Implement only the directly supported missing owner/transition/order after
this gate closes. Verify first idle frame and sustained idle after skip in
both orders, held-input resumption, normal landing bank, and existing scene/
body/fire behavior. Preserve original state rather than clamping every frame.
This is a focused investigation pass, not a requirement to decompile every
level before resuming implementation. Broader movement/contact/animation recovery
continues after its state and verification boundaries are made explicit.
## Verification and remaining work
The source guards for return and chain complete against the original identity;
the managed diagnostic compiles and records48 port cases. Its sources and traces
have SHA256 provenance under artifacts/jimmy-cutscene-return/port-trace/.
Original hashes and production-source identity are verified before publishing
this research checkpoint. Existing packages remain the earlier baseline.
The original skip-to-idle causal explanation remains unresolved; no full original
frame or new native Windows parity is claimed.
Final checkpoint verification: chain guard passes6 direct calls,11 instruction
sites,6 virtual slots,11 ranges and8 actual GAM records; return guard passes12
sites,12 slots and13 ranges. The48-case diagnostic's recorded source/output
hashes match.539 original files remain unchanged. A diff against a2e89c6 confirms
no changes under src/, tests/ or scripts/. These checks do not certify the missing
original skip-to-idle frame; that remains the next source-recovery task.
+106
View File
@@ -0,0 +1,106 @@
# Post-cutscene tilt: existing-port trace
2026-11-05. User confirms ordinary landing tilt also occurs in the original,
but Jimmy must remain upright after a cutscene. The earlier camera-entry fix
does not cover that resumed-idle boundary. Gameplay implementation is on hold
while the full original lifecycle is investigated.
## Scope and reproduction
This is **port runtime evidence**, not execution of original functions or a
decompilation claim. Baseline is `a2e89c6`, with unchanged production sources.
The read-only diagnostic at `tools/recon/cutscene-handoff-trace/` compiles the
existing Assets/Game sources into its own managed executable. It creates the
same synthetic flat-floor player as DirectorTests and runs either a synthetic
CAM or the actual supplied lab `neutron1a` MCA chain with a deterministic
two-second-per-speech clock. There is no GPU, SDL, full original scene collision,
or original executable execution in this diagnostic.
All48 cases complete successfully:
| Input/player order | Cases | Retilt after300 neutral ticks |
|---|---:|---:|
|Input before player update (fresh order)|24|24, approximately ±10 degrees|
|Input after player update (retained order)|24|0|
Each order covers both camera types, Left/Right bank, natural completion/Skip/
explicit DEFAULT, and jump/movement held or released during control lock. Each
case creates eight real controlled banked-jump ticks, triggers the camera,
waits until the body is grounded while control remains NULL, completes the
camera, then supplies **neutral** input for300 ticks. Held input during the lock
does not change the result because input/controller phases remain gated there.
The native camera source recovery belongs to
[camera orientation ownership](jimmy-cutscene-rotation-source.md), not this
managed probe. Its original addresses are not substituted for port execution.
## First stale writer
Representative actual lab MCA case: fresh order, Left bank, natural completion,
released controls during the scene:
| Boundary | Current Euler | Destination Euler | Mode / turn |
|---|---|---|---|
|Before trigger|(343.17676,1.3944445,357.05692)|(-40,1.2,-11.666668)|1 /11.666668|
|After camera preparation|(0,1.3944445,0)|(0,1.3944445,0)|1 /11.666668|
|Grounded during scene and after control returns|(0,1.3944445,0)|(0,1.3944445,0)|1 /11.666668|
|First neutral input publishes|(0,1.3944445,0)|(0,1.3944445,-10.000001)|1 /10.000001|
|End of first resumed player tick|(0,1.3944445,360)|(0,1.3944445,-10.000001)|0 /0|
|End of300 neutral ticks|(0,1.3944445,350.00018)|(0,1.3944445,-10.000001)|0 /0|
The first old-angle write is the resumed input phase, observed through the
existing `InputActions` event after `StepInput` publishes its target. It is not
an old current-Euler vector restored by the director. The director grants
control with the player still upright and facing clean.
In `LabSimulation.Step`, line279 dispatches fresh input before the player update
at line290. `StepInput` selects the air branch using the retained mode1 at
lines481–507, decays the old keyboard turn, and publishes the negative turn as
destination roll. Only later does the supported/released `StepPlayer` reset
mode/turn at lines555–560. Neutral mode0 input then retains the stale destination
because it has no accepted heading update (lines487–491). Consequently the
independent facing phase converges to that bank indefinitely. The first resumed
tick's intermediate360-degree representation is preserved in the trace; the
next tick visibly rotates toward the retained target.
With retained input-after-player ordering, `StepPlayer` resets mode/turn before
input dispatch. That input sees mode0 and retains the already upright target;
the bank never returns. This is a diagnostic distinction, **not evidence that
all scenes should be forced into retained ordering**.
## Evidence files and commands
`artifacts/jimmy-cutscene-return/port-trace/observations.json` retains the48 full
cases and selected boundary snapshots, including current/target/dirty, mode,
turn, speed, support, raw owner, current camera, root, controller and physics
counts. `summary.json` records per-order outcomes, baseline commit, source/harness
SHA256 identities and the observation-file hash. `run.log` records successful
completion. All are local generated outputs.
From GamePort:
```sh
source scripts/env.sh
dotnet run --project tools/recon/cutscene-handoff-trace/CutsceneHandoffTrace.csproj \
-p:NuGetAudit=false -- ../Original \
artifacts/jimmy-cutscene-return/port-trace/observations.json
```
No runtime, regression test, build script, original file or packaged binary is
changed. Parent owns integration and any eventual implementation decision.
## Implication and unresolved original questions
The confirmed port cause is a stale **air turn/movement mode** surviving camera
preparation, then being consumed by the first fresh neutral input. An additional
upright Euler write on DEFAULT alone would not explain or remove this first
writer unless its timing/state ownership also changed. The correct original
boundary must be established before implementing a reset or scheduling change.
Remaining questions belong to the independent original investigations: whether
source control return lets the player update run before input in that same
registry pass; whether camera activation/cleanup changes turn/mode through an
unrecovered callback; and whether the actual trigger/registry ordering differs
from the port's post-player interaction phase. This diagnostic does not answer
those questions or certify that the original returns with the same retained
state. Ordinary landing-idle tilt remains a separate confirmed behavior.
@@ -0,0 +1,99 @@
# Cutscene completion and first resumed input: original boundaries
2026-11-05. User comparison: ordinary landing tilt occurs in the original;
**skipping Carl's laboratory conversation must leave Jimmy upright**. Gameplay
changes are paused for source recovery. The previous camera-entry correction is
preserved; it did not verify neutral input after the scene ended.
Original EXE SHA256:
`f5281076e1a2492f2d61f87aae3f4544b6d0f535ebe33f9fbb6f78710fe1aea8`.
Retained REA full-instruction Evidence:
`ev_0d64cc58fa8f5573fc8e61e0f2c20e36288099212bd2257d3a3121361fd6c904`.
VA400000 is the preferred base; the guard uses PE sections for file offsets
(the cited text offsets equal VA minus400000). Originals are read-only.
## Recovered state owners
| Boundary / VA (file offset) | Original facts |
|---|---|
|Control118→475A40 (75A40)|Load argument, store raw controlled pointer5099E4, return. No actor, movement-mode, turn or orientation reset.|
|DEFAULT447D34..447FC5 (47D34..47FC5)|Restores raw owner through118; handles shared camera eye/UI/music and clears shared skip. Inspected code has no direct Euler/mode/turn reset or Jimmy Init.|
|MCA completion4311F0..43127F (311F0..3127F)|Resolves named PlayerControlled/NULL/none, grants through118, cleans resources and dispatches the next trigger. A grant is not initialization.|
|Jimmy cleanup42AE00..42AE29 (2AE00..2AE29)|Calls complete178, sets game1D65 and global4F8182. No direct movement-mode/turn reset.|
|Selected-gadget cleanup427FF0 (27FF0)|Raw-1 is rejected by the unsigned range guard428027; raw3's table entry also reaches common tail42882F. That tail only restores selected-gadget sentinel-1 and returns. Bubble/vehicle branches have separate policies.|
|Menu prelude405FB0/406590 (5FB0/6590)|The inspected selection/close callbacks manipulate menu resources. They do not dispatch a recovered Jimmy movement reset.|
|Player input438A60 (38A60)|Dispatches by retained game704, after raw-owner and camera-bump gates. Complete124 is ground438BC0;128 is air439900; no derived Jimmy air override supplies a hidden reset.|
|Air target439DB7..439E33 (39DB7..39E33)|Accepted pitch/current yaw/negative stored turn become destination Euler. This is a writer, not a render-only cache.|
|Grounded released controller437BA5/437BAC (37BA5/37BAC)|Clears mode704 and turn614 after the raw-ownership gate. Null ownership therefore leaves these fields retained.|
Jimmy complete object is game interface+C0. Game704/614 correspond to complete
7C4/6D4. Current and destination Euler retain their previously recovered separate
owners. Complete table4A3BDC slots124/128/12C/130/134/140 resolve438BC0/439900/
43A120/437C40/437F90/429910 respectively.
Input callback426030 has independent gates:407060 reads global4F8198,475CA0,
raw owner,4F8430 and game1D2D precede438A60. The bounded camera/default paths do
not establish a special first-resume write to4F8430 or1D2D. Known1D2D writes
belong to hit lock (complete1DED, set42A820/clear42567F). Camera bump7C8 can
suppress dispatch, but no inspected camera/default write proves it is seeded
here. These are ruled-out local explanations, not whole-program absence proofs.
## Skip is a distinct source path
The actual MCA update430DA0 tests a playing sound before entering its skip branch
430EED..431027. Escape509853, Space509858, controller509AC4 or shared skip4F83E4
can request it; nonzero4F8182 suppresses it. The accepted path resolves JIM1,
sets **complete1E00=1 at430F37**, sets **shared4F83E4=1 at430F3E**, stops the
sound, performs target-class/animation cleanup, deactivates the MCA, dispatches
completionE4 at431012, and resets its shot index. CAM has the analogous latch/
flag writes at41613B/416142 and its own completion order.
The Jimmy byte is the shared input-action/menu debounce latch recovered in
[gadget controls](gadget-controls.md), not a recovered orientation-reset flag.
No claim is made that copying this latch alone fixes the tilt. DEFAULT clears
shared skip at447D91 or447F89. The port's immediate SceneDirector.Skip/Drain
fast-forward does not represent these per-actor/registry/input-latch boundaries.
Its earlier comment labeling it an original chain skip overstates that coverage.
## New REA pseudocode and limits
A generated EXE copy was opened by local REA/Ghidra12.1.4. Complete replies,
Evidence export and closed snapshot are in
artifacts/jimmy-cutscene-return/rea-completion/. New successful pseudocode:
| Procedure | Evidence |
|---|---|
|475A40 raw owner setter|ev_3e6ebc517bdc52a7c2cd8af4475b837f2e7126f5a54a45e147356784b4e93eb7|
|438A60 retained-mode input dispatch|ev_77ee22e6d49adc0c983af92ba74273faf5f4daf697ab520ea45d22b08bc11bcb|
|42AE00 cleanup wrapper|ev_2ee50b10018363378445be32df38ab4f6b2125247b69485304867674d5d484da|
DEFAULT entry447CF0 did not resolve to a Ghidra procedure; the provider then
reported the numeric selector as an unmapped file offset. Its complete error is
retained as00.json; no pseudocode is attributed to it. The guarded retained REA
instructions support the DEFAULT statements above. No unchanged retry or claim
that the whole original was executed follows from that provider result.
## Port cause versus original explanation
[The port diagnostic](jimmy-cutscene-port-trace.md) reproduces48 cases. On fresh
input-before-player order, a grounded player returns with retained mode1/turn;
first neutral air input republishes the old bank before the controller clears
mode/turn. On retained order, the controller clears those fields first, avoiding
that writer. This identifies a port cause. It does **not** prove that changing
all scheduling or resetting turn on every control grant matches the original.
The complete original skip-to-idle frame has not been executed or fully closed.
Remaining explanations include exact skip/input suppression, actor callback
placement, contact/support state when the conversation starts and unmodeled
wrappers. User observation takes precedence over assuming the local component
composition describes the whole game. Further source recovery must explain why
the original avoids the stale writer before another runtime patch is made.
## Verification
`tools/binary/jimmy_cutscene_return.py` checks original identity,12 source sites,
12 virtual bindings and13 bounded spans. `jimmy_cutscene_chain.py` independently
checks skip/contact/frame boundaries and actual lab records. Complete generated
reports are under artifacts/jimmy-cutscene-return/source-contracts/.
These are static byte/data checks. The48-case harness is managed port execution;
no new original full-frame or native Windows observation is claimed.
+5
View File
@@ -1,5 +1,10 @@
# 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](jimmy-cutscene-decompilation.md) 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.