Files
Neutron/docs/decisions.md
T

2.6 KiB

Decisions

  1. Preserve input directory. Keep Original in place, baseline all file hashes, generate outputs elsewhere. No legacy binary execution in the source directory.
  2. Nested local repository. Workspace .git is reserved and not a functioning repository. Initialize GamePort/.git, no remotes, commit code/docs/tools only.
  3. Evidence-first data recovery. Native x86 executable does not imply little endian formats: GAM and OMT use big endian metadata. Require bounded readers, full consumption and real original-file regressions.
  4. Initial scene. NewGame.tsk starts LV1B and player GAM references STARTEXP. Prioritize that scene; use smaller VR01 for data/render verification. The LV1B LABEXP start trigger means simple camera spawning is insufficient for fidelity.
  5. Local dependencies. Missing SDK/native tools installed under .tools or /tmp after approved official downloads, without assuming root or altering host.
  6. Useful parallelism. Three independent agents own binary/rendering, asset recovery and infrastructure, while primary owns gameplay data and integration. Shared runtime/format changes must be coordinated. No external coordination.
  7. Honest compatibility. Build, headless Vulkan frame, SDL presentation, gameplay fidelity and native Windows execution are separate evidence categories.
  8. Preview ground contact and idle clarity. User testing exposed floor overlap and a misleading animated-default description. Original idle vertices are identical; do not invent idle motion. Expose moving source clips explicitly. Until original actor transforms/collision settling are recovered, apply one model-only floor correction derived from standing soles and visible geometry, with explicit preview labeling and regression tests. Preserve GAM/ASE data and per-frame morph displacement; this is not the original collision algorithm.
  9. Controllable lab before complete scene. Integrate evidence-backed no-gadget walking/jumping/camera now, with separate stateful simulation and explicit provisional source-triangle capsule collision. Rendering does not set the number of movement calls: use60Hz until original cadence is recovered. Preserve source GAM/ASE positions; keep physical feet, inferred actor-root height and visual pivot separate. Include invisible geometry rather than inventing visibility masks. Document the skipped intro, missing interactions, transient jump-state simplification, animation timing and close-wall camera limitation. Test against source geometry and build both packages without calling this a faithful full scene.