Files
Neutron/docs/checkpoint/implementation-plan.md
T

14 KiB
Raw Blame History

Evidence-based continuation

Checkpoint exit is architecture coverage, not full-game completion. Preserve baseline69d0187 and original hashes, finish reports/byte checks, then execute these chunks in dependency order. Each chunk gets its own evidence, regression result and local milestone; a failed gate triggers further bounded investigation rather than invented behavior or new levels/features.

Chunk Concrete change / evidence gate Verification and completion condition
C1 45F990: notify objectives before changing task records; then dispatch actors. Expose explicit pre-assignment boundary. Byte assertions45FB5B/7F/D9; observers see old task during objective selection and new task during actor notification; repeat/missing tasks still broadcast. Run Data/Interaction/Director/route regressions.
C2 Preserve per-actor virtual task override boundary. Generic activation then Carl override413F30 before advancing to next actor, rather than applying all Carl overrides after all actors. Confirm base/derived call bytes; check actor notification sees final per-actor state and authored marker; regression task50→60 conversation/control and lab route. Do not add other class overrides without evidence.
C3 Build original registry/construction/input/frame phase oracle for lab. Split pending motion/controller/effect phases only when registration order and engine callback bridge are recovered. Assert physics before normal object update, next-node timing, transform sync after callbacks. Compare input→impulse→next-physics→ground/pose trace; test pause/control loss/return. Keep cadence as unresolved policy until engine timer evidence.
C4 Recover47BD80 and Jimmy root/body fields. Replace provisional capsule/feet and wall response behind bounded source adapter. Depends on C3 phase semantics. Isolated original collision probes plus deterministic wall/step/ceiling/floor/dynamic-body cases; no invisible NC terrain; lab/tutorial route and visual root checks.
C5 Recover follow ray47C210 filtering and coefficients, then implement shared viewport handoff/DEFAULT snap and follow43B820. Source byte/probe predicate, below/equal/above1000 eye-distance boundary (snap strictly above), obstruction hit adjustment, controlled/scripted/return camera continuity, per-level projection. Port GPU checks remain separate from original visual comparison.
C6 Recover source allocation/failure and sound+44 duration conversion. Source slot14 uses software duration milliseconds and repeat+30, not hardware status. Model that channel state plus current SFX bank/default gate, then connect CAM/MCA completion. Strict elapsed>duration boundary, repeat/stop/invalid handle/clock pause; bank unload/reload/default/scene disposal. SDL dummy tests validate transport only; queue-empty is not source timing. Physical/native Windows output remains separate.
C7 Queued TSK load460E70→460EA0→4603F0: repeated task424 and all-entity420 post-task hooks. Ordinary scene placement separately uses46E61A→Jimmy40C423610. Recover enclosing loader, RETURN flags/root/STRT and acquired-reset gate50988A. C1/C2 first. Trace source duplicate-callback effect before adopting; original12-save byte round-trip; fresh load vs RETURN vs practice No/Yes; persistent arrays/score/energy/strings vs reset timers/handles/actors.
C8 Morph playback flags/rate/end events and UI dispatcher mode/layout gates. Original clip/end-state probes; action interruption/landing/cutscene restoration; objective/practice controls. Preserve modern widescreen/debug policy and test pointer anchors separately.

Implementation expansion into later levels, gadgets/combat, menus and full save UI resumes only after these core contracts support it. Later-level investigation is allowed where it resolves a core-system question, without waiting for every level.

Execution ledger

  • Entry: baseline69d0187 preserved;539 original hashes verified. Dedicated investigators assigned independent loop/motion, script/state and camera/audio/UI files. Main loop findings integrated by primary after investigator usage limit.

  • C1: implemented pre-assignment objective boundary. Data/Interaction/Director, input-only lab route and RETURN regressions pass. Source45FB5B/7F/D9 checked independently and retained in the script/state helper. No scheduler changes.

  • C2: captured SCENE before generic activation, moved LV1B Carl override into the same per-actor director callback, and removed the interactions-layer batch fix. Original413F3D/413F4E/414068..414118 independently checked. Director callback observers now see the final derived state before later actors. Director, Interaction and input-only lab route regressions pass; broader checks pending.

  • C3 investigation: input/frame bridge and fresh lab partial construction order recovered and byte-checked. Fresh game input→Jimmy controller→fire runs after global physics. Retained registry/lifecycle remains the integration gate; no speculative global reorder applied.

  • C6 investigation: voice-clock.md resolves derived slot14 to inherited software elapsed/duration and+30 to repeat. Replaced the proposed hardware/SDL-playing interpretation before changing runtime. Original bytes and one new regression pass.

  • C5 investigation: camera-ray-predicate.md recovers static grid registration, wall priority, FIFO first-hit and source kernel, with four order/data regressions. A static adapter is now justified; complete camera handoff still needs its state.

  • C7 investigation: scene-loader.md closes app deferred request→teardown→GAM→ per-object initialization/PostLoad. Jimmy/Goddard retention/reuse is proven; per-field reset/old-game/helper lifetime remains a bounded next trace. Editor R reset distinguished from ordinary load. Consistency review leaves old-game removal/eligibility explicit rather than treating current pointer overwrite as deletion. Five evidence regressions pass. This is the next C3/C7 integration gate.

  • Exit review: core architecture checkpoint documented in review.md; feature expansion remains paused. C1/C2 full integration passes; final follow-up Python discovery and original hashes recorded under artifacts/re-checkpoint/.

  • C3/C7 gate closed: normal Init copies stored+11 to live+139, restoring the level game's deletion predicate. Destructor explicitly unlinks its registry node. level-game-lifetime.md and three byte regressions retain that proof.

  • C3 first implementation: controlled player physics consumes prior motion/queued impulse before input/player/fire callbacks. Fresh versus retained input order follows proven registration; stored mode and same-frame landing reset are explicit. Twenty movement groups, input-only routes and GPU RETURN pass targeted checks. Full-scene scheduling, uncontrolled-body and collision response remain open.

  • C5 static adapter implemented/integrated for LV1B/VR04/LEV1: source cell/FIFO/ wall-first predicate and single-triangle kernel, source eye adjustment/gain, one query per input callback. Eleven camera groups and existing support groups pass; full camera handoff/input state remains open.

  • C7 availability correction: level entry enables state4→1 for the recovered eight gadgets, then VR01..08 disables owned1→4. Unknown states/other groups preserved; tutorial bypass unchanged. Source anchors and boundary/round-trip tests cover it. Arbitrary retained scalar/mode/fire state is still gated by the streamer/lifecycle fixture; no blanket carry was introduced.

  • Full integration passes: artifacts/implementation/integration-final.log. Seventy Python checks, all managed groups including20 movement/11 camera/ 13 support, GPU story/route/practice/RETURN/jetpack/animation with zero validation errors, SDL dummy lifecycle and539 unchanged hashes. RETURN also verifies retained input order and tutorial availability normalization. Packages rebuilt after the implementation milestone; platform runtime claims remain separate.

  • Packages: clean09cf117 Linux/Windows self-contained outputs; Linux launch from /tmp passes practice/RETURN and outside scene load. Release validation layers unavailable; source full integration validated above. Package evidence recorded separately without changing runtime. Continue the remaining source-state and shared-camera/control/collision gates rather than resuming feature expansion.

  • Retained-field continuation2026-10-26: property extraction has no hidden reset callback; generic Init re-applies authored body vectors, final preparation clears velocity/speed, Jimmy jump state resets separately from mode/hold/fire. Bounded route/no-gadget/jetpack PlayerLevelState now carries only those three scalars through ordinary transitions, never TSK loads. Companion/shadow reset fields and same-state AI guard are documented; body registration/new-helper timing remain.

  • Corrected C7 prefix-only availability: full VR01..08 tails re-enable the practice gadget0..7, so owned VR04 jetpack finishes1. LV1B disables gadgets1/7. Managed state boundaries/unknown values and GPU practice/RETURN guard the correction. Other level-specific selection/availability tails remain an explicit gate.

  • Retained continuation validation:84 Python,22 movement plus remaining managed groups,26 enabled-validation GPU renders, final nondefault tutorial/RETURN boundary replay, four final SDL dummy checks and539 original hashes pass. Original43A750/vector leaf component probe passes in isolated Wine32; this is separate from whole-loader/original-game/native-Windows validation. Next gates: uncontrolled body/controller/fire scheduling, body registration/helper traversal, C4 response/root and shared camera handoff. See progress.md for sampling-check details.

  • Release evidence: both self-contained packages from clean9ad5a8e; standalone Linux launch from /tmp verifies nondefault tutorial/RETURN scalar boundaries and body resets. Release validation unavailable; source GPU checks enabled. Native Windows runtime remains unverified for this chunk. Preserved baseline untouched.

  • C3 control continuation: gated movement/input independently of eligible body and fire; explicit CAM/MCA/AI raw-owner preparations and body-before-STOP order; initial MCA target-none returns, later shots do not repeat preparation. Scripted player transforms synchronize into body state. Grounded ordinary radius is outside the control gate. Full wrapper/base/ceiling/class effects remain open.

  • Same binding is exercised by the input-only lab route; modern replay now brakes before large turns instead of coasting past the narrow door. Task/score/inventory/ contact assertions remain. The observed corner behavior is retained as C4 input.

  • Local REA/Ghidra/full-JDK setup and batch Evidence/snapshot workflow verified. Five selected source procedures cross-check gates/base calls. Missing analyzed fire entry is recorded; no target execution or client registration change.

  • Final C3 integration: full scripts/test.sh passes with93 Python checks,24 movement groups, remaining managed checks and the binding-based input-only lab route, 27 zero-error enabled-validation GPU renders, SDL dummy lifecycle and539 hashes. Native GTX1070 Carl/control/F1/F3 commands pass; native Windows remains separate.

  • C3 release evidence: both self-contained packages identify clean c285b57; standalone Linux from /tmp passes airborne control/body/fire and nondefault tutorial/RETURN assertions. Release validation unavailable; enabled-validation checks above remain separate. Baseline untouched; Windows cross-published only. Logs: artifacts/uncontrolled-phases/package*.log. Next work is full wrapper/base ownership, C4 collision/root response and shared camera handoff.

  • C4 bounded kernel: REA/byte cross-checks recover47BD80 and helpers, with ten isolated original Wine32 synthetic cases and a managed numeric-reference module. Live capsule replacement remains gated on grid traversal/order, negative-Y split, normal callbacks, support/friction and dynamic reclipping. No source generic recursive capsule slide was found in the inspected dispatcher.

  • C3/C4 wrapper: root+150 upward query and min(−0.6Vy,−85) response now run before Jimmy movement/fire, regardless of ownership; mode1 cancellation preserves hold and independent queued/body work. Fresh/retained input boundaries preserved; camera-only debug switch cannot disable physical query. Other scene filters, sound and selected-gadget cleanup remain separate.

  • Registration: retained registry objects have49E cleared, then physical Init/list registration repeats Init before placement. Shared-eye capture/pre-update aim is source-guarded but C5 runtime correction remains next. Detailed reports: collision-response.md, ceiling-wrapper.md, body-root-registration.md, camera-handoff.md; complete REA results and provider gaps retained on disk.

  • Final C4 bounded continuation validation:104 Python plus HUD,25 movement/five sphere and remaining managed groups,27 zero-error enabled-validation GPU renders, SDL dummy lifecycle and539 unchanged originals. Four native GTX1070 window commands pass (five scene renders). Full capsule replacement, complete camera state and original whole-game/native Windows validation remain separate gates.

  • C4 continuation release: both self-contained packages identify clean57dfcc6 (source_dirty=no); standalone Linux from /tmp passes control/body/fire and nondefault tutorial/RETURN checks. Release validation unavailable, enabled source/native checks separate. Package logs in artifacts/collision-response/; baseline untouched and Windows cross-published only.