Files
Neutron/docs/progress.md
T

259 lines
17 KiB
Markdown

# Progress — 2026-10-25
**The full reconstruction is incomplete.** The current implementation extends the
user-verified Windows rendering checkpoint into an interactive original lab and
jet pack tutorial flow. This is not a verified faithful complete scene or release.
## User-requested live debug overlay
F2 toggles a compact, non-pausing live diagnostics panel. F1 also offers Live
Overlay and takes display priority while open. The setting/timing object belongs
to the session, preserving visibility across scene rebuilds and resets without
writing original saves. --debug-overlay enables startup visibility.
The12 rows expose FPS/frame time, root/velocity, steering/pose/ground/energy,
control/force/pause, tasks/score/jetpack, actual camera/far plane, active cutscene
and counters/errors. Source camera and orbit projection values reflect rendering.
The existing font/overlay buffers are reused with a smaller content-sized panel.
Targeted model/renderer/diagnostics groups pass. Native Linux GTX1070 SDL checks
pass live movement77.58 travel with visibility/menu/resize checks all true and
zero Vulkan validation errors (artifacts/live-debug-window.log). The prior F1
force-control check still passes. Headless GPU output was visually inspected.
Complete scripts/test.sh passes, including41 Python checks, all managed/gameplay/
GPU/audio regressions and539 unchanged originals (artifacts/live-debug-full-tests.log).
The lab -> VR04 -> lab headless check retains enabled diagnostics with frame
samples1/2/3 and zero Vulkan errors (artifacts/live-debug-return.log); exact RETURN
root/companion placement still passes. Final authored-camera inspection also
renders with zero errors, and targeted camera-mode/layout/timing checks pass.
Both Linux and Windows self-contained packages are rebuilt successfully. No new
Windows/Wine runtime check was run.
## User-requested F1 debug menu
Follow-up: Force Control is now a persistent TRUE/FALSE menu toggle. It overrides
effective ownership while preserving authored control state; switching it off
restores that state. Reset/new scene clears it. Cutscene execution continues and
DEFAULT music restoration uses authored ownership, avoiding a debug-induced silent
return. Menu/director/interaction checks pass; the extended16-frame Linux SDL
test records actual forced movement during active neutron1a,14.00 total travel,
all menu checks true and zero Vulkan validation errors
(artifacts/force-control-window.log).539 original hashes remain unchanged.
Both executables are republished for this follow-up; no Windows runtime
check is claimed. The full-suite evidence below is from the preceding menu build.
F1 opens a modern debug overlay with live player/scene/task diagnostics. Keyboard
actions cover camera, cutscene skip, energy refill, jet-pack grant, staged SCENE
editing,33 prepared levels/original starts, reset and resume. Gameplay, cutscene
clocks and SDL audio pause while open; input is captured through release after
closing. No new asset/dependency or original-game behavior is claimed.
Targeted menu/overlay and actual SDL pause tests pass. Native Linux GTX1070
16-frame SDL test passes pause, actions, Escape/held-input capture, resume and
800x600 resize with zero Vulkan validation errors (artifacts/debug-window.log).
Headless menu GPU readback was visually inspected. Complete scripts/test.sh passes,
including41 Python checks, all managed/gameplay/GPU/audio regressions and539
unchanged original hashes (artifacts/debug-menu-full-tests.log). Both platform
executables are republished successfully. No new Windows/Wine runtime check was run,
consistent with the user's earlier preference. See debug-menu.md.
## RETURN placement and underground passage
Recovered and integrated the next lab checkpoint:
- TSK StoredVector is the departure root at Jimmy+9A8. Ordinary transitions
capture the live root; RETURN preserves it, restores Jimmy and Goddard to it,
retains destination STRT yaw and suppresses the initial STRT trigger once.
Fresh TSK loading does not imply pending RETURN. Pickup42AF00's separate
+1E38/+1E48 checkpoint remains unconsumed rather than being conflated with it.
- Original shared level loaders omit NC-prefixed SHAPE names from terrain
registration. Both player terrain and source actor support now use that rule.
All2869 lab render/source triangles remain;2851 are collision-eligible, including
all325 null-material proxy triangles. No material-name water filter was added.
- The underground replay starts at the original riverlab STRT with explicit
SCENE65 fixture, then uses only live movement/contacts. It operates both switches,
crosses fully clear bars twice, waits for reclosure and checks the closed box
blocks a source-gated negative sweep.2763 steps,4153.21 travel, no jumps/warps.
The former85.89-unit first-step rise onto decorative water is now rejected.
The headless lab -> VR04 -> lab RETURN check restores the actual departure root
(4921.8735,76.49398,502.37677), separate from STARTEXP, with no intro playback.
Complete scripts/test.sh passes:41 Python checks,18 movement groups,11 collision
groups, four source/checkpoint groups, director/route/underground regressions,
headless Vulkan scene/RETURN checks, audio lifecycle and539 unchanged hashes.
Evidence: artifacts/return-underground-tests.log. The three-scene RETURN check
includes exact companion placement and reports zero Vulkan validation errors.
Both self-contained executables are republished successfully; no new Wine/native Windows
runtime check is claimed, following the user's preference to skip that check.
Reports: analysis/runtime-checkpoint.md, water-collision.md, underground-route.md.
The underground fixture does not establish progression into SCENE65; full water
hazards, Goddard follow/support phase, fades and original runtime comparison remain
incomplete. Native Windows testing of these new changes is not agent-verified.
## User-reported lab regressions (preceding milestone ca08c46)
The user confirms lab -> tutorial -> lab -> outside is reachable. Four follow-up
bugs are now implemented with source evidence and targeted regressions:
- Carl's neutron1a MCA no longer grants its JIM1 completion target at activation.
Active shots clear control; natural/skip completion restores it. Runtime holds
move/turn/jump for 60 frames and records zero player simulation steps.
- Grounded/released jump resets the source turn field (interface+614=complete+6D4).
Residual air turn previously opposed subsequent ground steering. 17 movement
groups pass, including landing into both opposite directions. All five jump
bindings pass identical SDL powered replays; no button-specific turn scale.
- Gameplay clipping now uses recovered original level overrides: outside28000,
tutorials50000, all35 level mappings. The old universal3200 clipped valid distant
scenery; actual outside GPU before/after readbacks confirm its recovery.
- Objectives have source 0.7-second entry, hold until 7, exit until 8.7. O opens/closes
via an edge-latched timer instead of changing layout. Modal/control gating and
source sound216 are integrated; separate pad/content draw checks pass.
The complete scripts/test.sh run passes, including 39 Python checks, managed
regressions, scene/GPU/audio checks and all 539 unchanged original hashes
(artifacts/four-bugs-full-tests.log). Both self-contained packages are rebuilt.
Packaged Linux popup/O/resize and 60-frame Carl control checks pass on GTX1070
with zero Vulkan validation errors. The already-started preliminary Wine64 popup
check also completed successfully (validation disabled). Further package checks
were skipped at the user's request: the user reports that their own Windows build
works with these corrections. That report establishes user verification, not an
agent-run native Windows comparison or complete input/audio/driver coverage.
Reports: analysis/carl-cutscene-control.md, air-steering.md, camera-clipping.md and
objective-timing.md. Original comparison, per-layer rendering/dynamic fog and full gameplay
remain incomplete. User loop reachability is distinct from source-faithful
checkpoint-position restoration and agent native Windows validation.
## Jetpack steering correction
User testing exposed frozen airborne steering. The earlier recovery incorrectly
read mode1's log branch as an early return; it actually falls through to439900.
The port now uses the original airborne Up acceleration, Down braking and
Left/Right turn-rate rules for jumps and jet pack bursts. Source byte checks,
16 movement groups,14 vertical burst groups and the full input-only lab route pass.
VR04 GPU replay travels403.88 horizontally during powered flight, turns and lands
with zero validation errors. The rebuilt Linux package and Windows package under isolated Wine64 pass300
SDL-input frames with Up/Left/Down/Shift and resize800x600, matching403.88 travel.
Linux GTX1070 validation reports zero errors; Wine validation is disabled.
Native Windows retesting remains unverified by the agent. Details and remaining fidelity limits: analysis/air-steering.md.
## Latest user-tested correction
User testing found background music continuing during cutscenes. Recovered CAM
416776..4167E1 stops/unloads the selected music; DEFAULT447F12..447F45 restarts it
when control returns. The runtime now performs that handoff. Source-byte/director
checks and actual SDL dummy streams verify silence through all16 lab cues, normal
completion restart, reset silence and skip restart. See analysis/sound-effects.md.
Both platform executables were republished with this correction. The packaged
Linux and isolated Wine64 Windows executables pass the full music-lifecycle audio
check. Linux packaged prompt/input/resize still passes with zero Vulkan validation
errors. Native Windows listening remains unverified by the agent; source user
report and exact static rules support this fix.
An additional managed replay now walks through explanation, jet pack, Carl and
outside LOAD using only movement inputs, source live contacts and NPC patrol/body
updates. The current replay reachesSCENE60/score150 with jet pack owned after887 movement steps
and7078.63 travel; the player is never repositioned. This strengthens port lab
navigation coverage without proving original collision/cadence fidelity.
## Current integrated behavior
- Default play runs LV1B STARTEXP and its16-cue CAM/MCA/AI camera/speech sequence.
Escape/Space skips actions and restores control; `--inspect` retains inspection.
- Original walking/backpedal/turn/jump arithmetic, source clips, chase camera and
fixed60Hz port simulation. Powered jumping now translates the recovered state
table,8-second effect timer,60/100 contact radii and energy/VR immunity rules.
- Original Carl/Goddard morphs, sprite quads and9 lab model actors, including the
OMT phones, source buttons and moving bars. Goddard's unregistered STOP/WALK/TALK
requests retain the current source clip rather than invented aliases.
- Shared patrol states1/3/7, state9 base motion and Carl's WALK/STOP profile.
Body support uses recovered radii, factors, quick magnitude and one-sided trace;
full wall response and Carl sensing remain incomplete.
- Lab explanation -> SCENE35 -> jet pack pickup -> SCENE50 -> Carl contact/talk
-> SCENE60 -> outside LOAD dispatch. Pickup adds50, Carl callback adds100 before
dialogue. Source task gates/objective selector and inventory slot0/3 persist.
- The pickup's internal tag `shrinkray` names the lab jet pack, proven by original
GET JET PACK art and powered-jump inventory checks. It is not a shrink-ray weapon.
- Original objective pad, three lab objective images, digits and PracticeYesNo
are rendered at source640x480 coordinates. O opens/closes the timed popup.
Practice pauses the simulation; Y enters VR04/PHONEBOOTH and N resumes.
Return context is current GAM/latest STRT (LV1B/STARTEXP for NewGame).
RETURN restores the saved departure root; the pickup checkpoint is separate.
- Button contact/cooldown/down/up/next-trigger and vertical-door duration,
reversal/hold/next-trigger logic. Source dynamic-box broadphase/triangle ordering
is translated; narrowphase/response still uses a modern approximation.
- Ordinary stars/trophy, score and acquired flags; all12 TSK files round-trip.
CLI checkpoint load/write preserves unknown fields, inventory, tasks, energy,
objective selector and RETURN context. Complete runtime/menu saving is unfinished.
- Music and speech use independent SDL streams. Positional effects/banks and
physical audio verification remain unfinished. Original default-bank pickup/button/
moving-door one-shots and owner loop stop are now connected; attenuation/volume
policy and jetpack sound callbacks remain unfinished.
## Source availability and evidence
All539 original hashes remain the baseline. All35 GAM files are supplied and33
have matching OMT geometry. Level3B/E GAMs exist, each1777 bytes, and are identical
after their four-byte level identifiers. Their classes name absent OMT archives,
but no serialized incoming GAM transition or supplied save selects them; loader
null handling does not establish a fatal failure. Normal gameplay use is unverified,
so these are not an established blocker and no extra files are currently requested.
The catalog preserves absent geometry without substitution.
Detailed recovery: analysis/lab-progression.md, scene-entities.md, scene-models.md,
patrol.md, original-collision.md, dynamic-collision.md, buttons-doors.md, ui.md,
jetpack.md, level-resource-mapping.md, and save-format.md. Addresses, source offsets,
confidence and approximation boundaries are retained there.
## Verification
Managed checks cover source data, director/actions/cameras, task/interactions,
patrol, recovered support, moving-box collision, jetpack states and HUD composition.
Python currently has39 top-level regression tests plus2 targeted UI extraction tests.
Existing11 collision and17 movement groups pass;14 jetpack groups pass. Eight
quick-magnitude float results are checked against an isolated original DLL probe.
Original mode2 standing/jump animation probes confirm zero dynamic-point offsets
for the tested clips. This is DLL evidence, not observed original game behavior.
Headless llvmpipe runs render the story, model actors, practice prompt, lab exit
and VR04 entry with zero Vulkan validation errors. The300-step VR04 powered replay
jumps once, rises418.98 units and lands; VR immunity retains20 energy. The corrected ordinary
240-step replay travels1264.55 and rises137.62; port landing-count telemetry changed
with the explicit state table. These measurements verify port behavior only.
See artifacts/full-integration-test.log and practice/jetpack logs for current checks.
The full scripts/test.sh integration passed at c929f94:35 Python regressions,2 targeted UI
checks, all managed groups, GPU story/route/practice/VR04 replays, SDL dummy streams
and539 unchanged original hashes. The subsequent music correction passes targeted
source-byte/director/SDL audio checks on both packages, plus native Linux window
and input-only managed navigation regressions; hashes were reverified.
Both self-contained packages rebuilt successfully.
Linux package runs the12s story without SDK/source/Original mounted, zero Vulkan
validation errors. Its GTX1070/Xvfb practice window pauses, receives N, resumes11
steps and resizes800x600 with zero validation errors and original pickup audio.
The Windows package passes the same12-frame practice/input/resize/stream check
under isolated Wine64, matching simulation statistics (validation disabled).
It also renders the12s story under Wine. Generated package logs live under
artifacts/package-integration.log, linux-package-integration/, package-practice-window.log
and windows-practice-package/. Initial native Windows loading/rendering remains
user-confirmed for the earlier checkpoint; current native Windows story/input/audio
is still unverified. Wine is preliminary compatibility evidence only.
## Next work and limits
No usable original gameplay reference has been captured: isolated Wine GL/Vulkan
attempts fail rendering. Continue static recovery and isolated DLL probes locally;
comparison on a working original installation remains required for fidelity.
Next priorities are progression from outside into later lab task states;
original collision47BD80/479D20 and phase/cadence; audio-bank/SFX mapping;
practice highlights/energy gauges; actor sensing/animation completion. Then expand
remaining gadgets, damage/combat/enemies, menus/configuration, full saves and all
level progression. Lighting/fog/blend/culling and sprite/pivot fidelity are open.
Do not claim complete gameplay based on direct-contact route or GPU smoke tests.
All previous agent file ownership has returned to primary; investigators that
reached usage limits are not dependencies. Generated output stays separate from
Original and the nested local repository has no remotes. Read tasks.md and Git
status before continuing; preserve source evidence and update these notes.