Files
Neutron/docs/checkpoint/review.md
T

6.4 KiB

Checkpoint exit and implementation continuation

The initial architecture checkpoint is complete: all requested core systems have source anchors, state owners/lifetimes, route interactions, baseline comparison, explicit unknowns and bounded implementation actions. Feature expansion remains paused. This is an architecture milestone, not complete original-game fidelity.

Exit criteria and evidence

Criterion Durable result
Entire specification/current findings/code reviewed Full1281-line prompt.MD, both AGENTS files, runtime and focused prior findings reviewed; stale architecture claims reconciled through checkpoint index
Preserve playable port and originals Clean69d0187 local baseline tag and independent Linux/Windows package copies;539 original hashes unchanged at entry and final verification
Main loop/input/motion/jump/jetpack/collision/animation loop-motion.md, input-frame-bridge.md, lab-registry.md; source frame order, two clocks, pointer interfaces, controller/effect ownership, pending collision/morph gaps
Scripts/tasks/cutscene/control/progression script-state-route.md; synchronous objective→records→per-actor callbacks, raw control ownership and route-specific script/task transitions
Camera/clipping/audio/UI camera-audio-ui.md, camera-ray-predicate.md, voice-clock.md; shared viewport, strict snap, first-hit grid order, software audio clock/bank lifecycle, modal/objective timers and modern layout policy
Level transitions/persistence/checkpoints/save-load scene-loader.md and script-state-route.md; deferred ordinary load vs TSK load, selective retention/reuse, placement vs post-task callbacks, distinct pickup capture/RETURN root
First route case study lab STARTEXP/intro→explanation35→pickup50/practice→VR04→RETURN lab→Carl60/dialogue→level1.gam/level1; state/owner/reset boundaries recorded
Coverage/assumption audit and concrete plan coverage.md and implementation-plan.md; source facts, inference, port behavior and verification separated; explicit per-chunk gates rather than waiting for every level

Three independent investigators used exclusive files; primary reviewed decisive bytes and integrated reports. All ownership has returned. No new levels, gadgets, combat or gameplay features were added. No originals were modified/executed in place. Existing remote configuration was left unused; work stayed local.

Implemented and verified chunks

C1: TaskStates.Assigning selects objectives before records change. Entity callbacks still follow mutation, including repeated and missing-name assignments. Tests assert old-state observation at objective notification and new-state observation afterward. This corrects original45FB5B/45FB7F/45FBD9 ordering.

C2: Carl's LV1B override now belongs to his director task callback: capture SCENE, apply generic activation, then marker/state/visibility override before the next actor. Removed the interactions-layer post-batch fix and redundant level-code argument. Tests cover callback observation, source activation boundaries and existing lab/Carl/RETURN routes. Other class overrides remain unported.

artifacts/re-checkpoint/integration.log: full scripts/test.sh passed, including 49 Python checks at its start, all managed groups, GPU story/practice/return/Carl/ objective/jetpack/animation replays, SDL dummy audio and539 original hashes. All GPU runs report zero Vulkan validation errors. Additional follow-up byte tests were added afterward; final Python discovery passes62 tests, recorded separately in artifacts/re-checkpoint/python-final.log. Build succeeds without warnings. These are port/static checks; no new native Windows runtime or original whole-game runtime verification is claimed. Physical audio and original cadence remain open.

Self-contained Linux/Windows packages were rebuilt by scripts/package.sh from clean531883a; both BUILD-INFO records identify that source and source_dirty=no. The preserved baseline DLL hash remains distinct from the new binary. Linux release launch from /tmp passes the60-frame Carl cutscene held-input lock check (linux-package-carl.log); validation layers were unavailable in that package invocation, unlike the enabled full integration runs. Subsequent changes only refine loader evidence and this validation record, not runtime code.

Final loader consistency review verifies that the lab predicate-zero writer is in class-default initialization44D3D0, reached from constructor44F3B3 through 4749A0→completeB4→44F4B0. It is not inside base constructor44D2B0. Five targeted loader regressions pass after adding that provenance. This strengthens the old-game lifetime question rather than silently assuming pointer overwrite deletes it.

Plan continuation and exact next gates

C3 investigation recovered fresh game input→Jimmy→fire order after physics. C5 recovered source static obstruction, C6 corrected the proposed audio status model, and C7 recovered the ordinary loader. These are executed investigation chunks, not implemented scheduler/camera/audio/loader replacements.

  1. Resolve old level-game registry removal/eligibility, since constructor overwrite of509948 is not destruction and its initialization can set the teardown predicate0. Trace retained Jimmy/fire/shadow/Goddard field resets through property read and per-object28/40C. Then construct the route's callback trace and split pending physics/controller/effect/sync phases. Do not impose fresh insertion order on retained objects.
  2. Independently implement/test the static camera query using source shape/polygon IDs, registration filters, sparse cells and FIFO wall-first reduction. Then integrate shared viewport handoff and follow state, preserving strict>1000 snap. Modern nearest-sphere clipping cannot stand in for that source algorithm.
  3. Recover sound duration conversion and allocation/play-failure cases, model software channel duration/repeat/stop/clock pause plus bank ownership, then change CAM/MCA completion. SDL queued bytes are transport state, not source time.
  4. Continue collision response/root, morph end events and UI mode/layout recovery under the remaining plan gates. Original runtime comparison is still required before claiming source fidelity, but does not block directly evidenced chunks.

Use implementation-plan.md as the execution ledger. Continue locally and autonomously; user confirmation is not a gate for the authorized investigations or directly evidenced corrections.