Fix Rocketship wall damage flash misinterpreted as model scaling

This commit is contained in:
UnixNight committed 2026-11-03 16:53:56 -05:00
1 parent b5ef5b8b64
commit 62e908e26d
13 files changed
+255 -21

No files matched your search

+98
View File
@@ -0,0 +1,98 @@
# Rocketship wall-impact render correction
2026-11-03, user-reported wall-hit distortion. **Recovered fact: the four
impact values are material diffuse RGBA, not geometry scale.** The earlier
Rocketship lifecycle report, source-projection naming and supplied friend's
`C3DRocketShip.md` called these values scale; that semantic interpretation was
wrong. The copied-original numeric observations remain valid.
## Ownership and source chain
Authoritative EXE SHA256:
`f5281076e1a2492f2d61f87aae3f4544b6d0f535ebe33f9fbb6f78710fe1aea8`.
Authoritative OMT2.dll SHA256:
`3ec41ead011adf56387a6dbd4580e5b5a9721b5f8020c8a56420f42824354943`.
The friend's different DLL is not used. Analysis reads generated copies only.
1. EXE40DCA0 allocates0x8C bytes;40DCC0 calls IAT48D18C, whose imported name
is `OMedia3DMaterial::OMedia3DMaterial()`.40DCCA stores this pointer in
`complete+5C0+textureIndex*4`.40DCD1/40DCD4 initialize its blend fields30/34.
This is the inherited indexed material array, not an object transform.
2. Rocket43E76E loads `game-interface+500`, equal to `complete+5C0`, material0.
43E780 adds4C. Stores43E78B,43E795,43E7A4,43E7A7 write offsets4C,50,54,58
respectively as `(1,pulse,pulse,1)`. No passenger, Goddard or flame pointer
is traversed by these writes.
3. DLL exported `OMedia3DMaterial::set_color` at10001320 writes the three RGB
components into offsets4C/50/54 (among its other material colors), proving
the fields belong to a color. Its name comes from our DLL export table.
4. DLL exported `OMediaDXRenderPort::draw_shape` at100374B0 consumes the same
material fields.100376A6 passes material+4C to render-port set_material
as its diffuse color. The unlit branch1003787C..100378F6 reads+58 first,
then+4C,+50,+54, multiplies components by255 and packs ARGB diffuse.
Thus offsets4C/50/54 are RGB and58 is alpha. Geometric vertices are not
scaled by this code.
5. Existing REA40E4A0 independently writes `(1,1,1,alpha)` at these same
material offsets while changing blend state. Existing Shrink Ray and
Invisible evidence already identified this as opacity rather than scale.
EXE text raw offsets above are VA−400000; the cited DLL code ranges have raw
offsets VA−10000000. Full bytes/hash guards are retained in
`artifacts/rocket-impact-render/material-byte-guards.json`:
| Source range, exclusive end | Raw start | SHA256 |
| --- | --- | --- |
| EXE43E76E..43E7AA |3E76E|52fb46c6a2271fd06691be976a277a53350aaf6dd2b7083a623ffa797eb8819f|
| EXE40DCA0..40DCF6 |DCA0|fd3810ec3db075e35a81025c0ebec5bd0c2504072306dafac78e5a818c817aff|
| DLL10001320..1000139B |1320|071ee5efa77f876b08994ddc3750a4b8e6f2581b741f1e4c8202285b12ff6a65|
| DLL10037694..100376B3 |37694|d9784e68b81bf68a5ef9df686fc4535675915f21643898bcda0dc1ade43562d5|
| DLL10037874..100378FA |37874|32a2fea66ca6438d44061ba85730509d8c7509ac045001688273c65042619ae4|
## REA evidence and probe interpretation
New local REA/Ghidra queries of the generated authoritative DLL copy succeeded;
complete replies, export bundle and closed snapshot are under
`artifacts/rocket-impact-render/rea-material/`:
- 10001320 set_color: `ev_71e12ebb8e6f6994553fe33606a358b7cd5a0277e50f8190b022e28216ced8ef`.
- 10036F70 DX set_material: `ev_991aa57164839ac125a6cfd9ea309896ef8e525037debed8aeb7f413da71c72c`.
- 100619C0 pipeline set_material: `ev_50a751ce5c2dae8b884f31c0f4da52b6f792054c9d30b74fa64739c84ffa1a76`.
Prior40E4A0 evidence:
`ev_615030aa60b2346083853c40aa719ebc466e9f393fd870ea9722536a4f64cfac`,
`artifacts/invisible/rea-invisible-ipc/01.json`. The DX set_material decompiler
loses stack-copy detail; exact draw_shape disassembly supplies the downstream
packed-color ordering. This is static shipped-artifact evidence, not observed
original gameplay.
The eight earlier impact-tail probes stubbed the material object and printed
the four writes under the misleading output label `scale`. They measured
values, timing, strict expiry and impulses; they did not execute a renderer
and therefore never established scaling. Preserve the old raw fixture field
for provenance, but interpret the observations as color.
## Implementation implication and verification boundary
Apply `(1,pulse,pulse,1)` only to the ship's diffuse tint, retain unit geometric
scale, and restore white when the source impact interval expires. Keep the
existing source pulse and strict timer comparison. The result is a red/white
damage flash with opaque alpha, not a squash/stretch animation. Passenger,
Goddard and fire transforms/tints remain independent.
The binary investigation performs no target execution. Primary integrated
ImpactColor(Vector4), ship-only vertex RGB/opaque alpha and unit model geometry.
Python projections now name the field color; future probes print rgba while
retained raw observations keep their old label. Three independent byte spans
guard the writer, material constructor and DLL diffuse consumer in regression.
The13-group managed suite passes. Its new regression drives a real static-wall
callback twice, samples all four original resource layers over130 frames, checks
every vertex against unit geometry, checks ship-only red tint and verifies
re-rendering does not accumulate color. It also verifies expiry and reset restore
white.12 lifecycle Python groups pass, including all eight numeric references.
Main .NET build passes. The existing1920×1080 ten-phase headless flight replay
passes with validation enabled and zero Vulkan errors; that replay has no wall
impact and is not a GPU wall-hit comparison. All539 original hashes verify.
Logs: artifacts/rocket-impact-render/{managed.log,python-material.log,build.log,
replay.log,originals.log}. Package verification is recorded below after rebuild.
Full original wall-hit rendering and native Windows comparison remain separate
evidence categories.