jan/rrt
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Preserve engine type profile cohorts

Browse source
This commit is contained in:
Jan Petykiewicz 2026-04-21 23:24:49 -07:00
commit f67629069c
3 changed files with 71 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

3
docs/rehost-queue.md
View file

@ -12,11 +12,12 @@ This file is the short active queue for the current runtime and reverse-engineer
- The active static parser head is now the `engine_types` semantics frontier.
The repo now has structural inspectors for `.car`, `.lco`, `.cgo`, and `.cct`, but the binary side is still only partially semantic: the checked 1.05 corpus grounds `.car` fixed strings at `0x0c / 0x48 / 0x84` plus a second fixed stem slot at `0xa2` and a side-view resource name at `0xc0`, while `.lco` carries a stable primary stem at `0x04` and only conditional companion/body slots at `0x0c` and `0x12` when the leading stem slot is padded.
The checked 1.05 corpus now also splits `.car` auxiliary stems into `126` direct matches, `14` role-neutral roots, and only `5` truly distinct cases, while `.cgo` collapses into five stable scalar ladders instead of arbitrary floats.
The checked 1.05 corpus now also splits `.car` auxiliary stems into `126` direct matches, `14` role-neutral roots, and only `5` truly distinct cases, with those five exact internal-to-auxiliary pairs now preserved directly in the report surface, while `.cgo` collapses into five stable scalar ladders instead of arbitrary floats.
The early `.lco` lane block is now partially partitioned too: only offsets `0x20`, `0x34`, `0x38`, `0x3c`, `0x44`, `0x48`, and `0x54` behave like low-cardinality buckets, while the other early lanes still look high-variance.
The side-view resource path is now grounded into `Data/2D/rt3_2IMB.PK4`, and the `.imb` parser now decodes shipped comment-suffixed numeric rows plus `_NE` profile fields such as `HorizontalScaleModifier` and `ImageWHScaled`.
The checked PK4 linkage split is now explicit too: `132 / 145` side-view resource names resolve directly, but the remaining `13` are the missing `CarSideView_3.imb` cohort and that hole exists in both checked installs, while `43 / 145` derived `{internal_stem}_NE.imb` names resolve and all of those hits belong to matched locomotive pairs.
The packaged profile metadata is stable enough to summarize: `CarSideView_1` is `512x512` at `0.04` VRAM, `CarSideView_2` is `512x256` at `0.02`, and every packaged `_NE` profile is `512x128` with `HorizontalScaleModifier = 0.75` and `MaxPercentOfInterfaceVRAM = 0.09`.
The `_NE` split is now aligned with the locomotive display census too: all `43` packaged `_NE` hits live inside the grounded display prefix, and all `5` unmatched display-tail families are still missing packaged `_NE` profiles.
The next honest static work is to keep promoting those fixed lanes into stable parser fields, explain the five remaining distinct auxiliary-stem cases, and decide how far the `.cgo` ladders plus the low-cardinality `.lco` lanes can be grounded without overclaiming semantics.
Preserved checked parser detail now lives in [EngineTypes parser semantics](rehost-queue/engine-types-parser-semantics-2026-04-21.md).
Preserved checked format inventory detail now lives in [RT3 format inventory](rehost-queue/format-inventory-2026-04-21.md).

8
docs/rehost-queue/engine-types-parser-semantics-2026-04-21.md
View file

@ -71,6 +71,12 @@ first `.car` / `.lco` / `.cgo` / `.cct` inspector pass landed.
- `CarSideView_2.imb`: `512x256`, `MaxPercentOfInterfaceVRAM = 0.02`
- every packaged `_NE.imb` profile: `512x128`, `HorizontalScaleModifier = 0.75`,
`MaxPercentOfInterfaceVRAM = 0.09`, and `ImageWHScaled` present
- The `_NE` profile split is now tied back to the locomotive display census:
- `43` matched locomotive display-prefix families resolve `{internal_stem}_NE.imb`
- `18` matched display-prefix families still do not resolve `{internal_stem}_NE.imb`
- `0` unmatched display-tail families resolve `{internal_stem}_NE.imb`
- all `5` unmatched display-tail families (`242 A1`, `Class 460`, `Class A1`, `Class P8`,
`Class QJ`) are missing packaged `_NE` profiles
## What The Current Parser Now Owns
@ -81,6 +87,7 @@ first `.car` / `.lco` / `.cgo` / `.cct` inspector pass landed.
- auxiliary stem slot
- side-view resource name
- auxiliary-stem relation counts across the shipped corpus
- exact distinct auxiliary-stem pair counts
- PK4-backed side-view resource resolution status
- derived `{internal_stem}_NE.imb` resolution status
- `.lco`
@ -102,6 +109,7 @@ first `.car` / `.lco` / `.cgo` / `.cct` inspector pass landed.
- `rt3_2IMB.PK4` link surface
- packaged side-view profile summaries for `CarSideView_*`
- packaged `_NE.imb` texture-size and scalar-family counts
- packaged `_NE.imb` match counts split by grounded display prefix vs unmatched display tail
## Remaining Static Questions