Files
Neutron/docs/checkpoint/dynamic-reclip.md
T

154 lines
10 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Populated dynamic reclip and recovered rounding boundaries
2026-11-07. Continues [dynamic context](dynamic-context.md). The complete original
box query now runs against a populated synthetic static grid. Fourteen full
observations and12 diagnostic kernel steps match the C# composition exactly as
parsed float32 values. This checkpoint corrects directly evidenced shared math;
it does not admit dynamic pairs or new actors to live gameplay.
## Evidence and runtime boundary
Canonical EXE SHA-256:
`f5281076e1a2492f2d61f87aae3f4544b6d0f535ebe33f9fbb6f78710fe1aea8`.
The generator reuses the prior11 guarded bodies/102 relocations, including
47A5E0,47B4B0,47BBB0,47BD80,469880 and469B70. Body hashes, PE offsets and copied
OMT2.dll identity are retained in `artifacts/dynamic-reclip/native/source.json`.
REA evidence is reused, not regenerated for an unchanged question:
| Evidence | Source |
|---|---|
|`ev_0d64cc58fa8f5573fc8e61e0f2c20e36288099212bd2257d3a3121361fd6c904`|Complete original instruction listing; decisive float stores/call order.|
|`ev_76827fca086933b7c48a9036a5a4a0a0f86d714935de6373eca4a3530fd231c5`|47A5E0 mode2 pseudocode, retained under dynamic-context/rea.|
|`ev_f39128dafc2a3a20b5768613ce22b3c3168fe8653908837d92ad9a69ef5f2ad8`|47B4B0 static traversal/retry; see collision-grid-traversal.md.|
|`ev_5e39f2668d04b1fdd2c91b298a28413327dc3879e6b8bbffe0a14bc1dcbc1128`|4695B0 nearest-edge pseudocode; assembly resolves intermediate stores.|
`tools/binary/dynamic_reclip.py` generates a bounded harness containing copies of
the existing probe sources without their mains. The new executable calls copied
source bodies with synthetic box/sphere getters, point transforms, callbacks and
grid nodes. It never invokes the original entry point. An instrumented private
mapped47BD80 counts primitive mode and captures12 steps only for the skew case;
the hook restores original entry bytes for each real kernel invocation. Original
files and on-disk copied EXE remain unpatched. This is Wine32 component execution,
not full original gameplay, live actor callbacks or native Windows verification.
Final execution is under `artifacts/dynamic-reclip/probe-verified/`, with no
network or Original-directory mount. An earlier overlapping prefix preparation
produced no observations; a separate completed prefix then exceeded the12-second
startup bound while still running. The45-second bound completed successfully.
Incomplete attempts and intermediate comparisons are preserved separately and
are not counted as source observations. One initial ceiling triangle missed the
intended path; the final fixture covers it and proves a real370 callback.
## Shared state through populated static geometry
The box is identity-oriented with cube(−50,−50,−10)..(50,50,10), radius40;
sphere radius10, previous(0,0,50), current(0,0,15). The dynamic face query normally
gives rootZ20, face40, edge10 and hitZ10. A separate old polygon99 is seeded.
| Fixture | Original result / dependency |
|---|---|
|Static wallZ35|Nearer static face wins: rootZ45, face15, hitZ35, last0;2 passes/4 cell-list visits.|
|WallZ10 at equal dynamic face|No replacement: rootZ20, face40, old polygon99 survives.|
|Far wallZ0|No replacement; old polygon99 still supplies the terminal callback.|
|Walls35→30 and30→35|Both end atZ45/face15; winning polygon is0 or1 according to its identity.2 passes/8 visits.|
|Root-mutating callbacks|Final local root still overwrites the mutation, now toZ45.|
|Box response inhibited|Only dynamic callback; no static visits, passes or setter.|
|Static geometry disabled|2 node visits, no static kernel calls; retained old-polygon callback still occurs.|
|Sphere42C enabled|12 dynamic kernel calls in mode0,4 static kernel calls in mode1, terminal callbacks/setter in mode0.|
|Old ceiling polygon|No accepted static face, but retained normal−Y invokes370 rather than374.|
|Real ceilingY35|Root(0,25,23.7499962), face43.7501259, last1;370 follows dynamic374, then314;2 passes/12 visits.|
|Repeated query without context reset|Nearest limits reset per dynamic query; last polygon and all callback/pass/visit/kernel counters persist across calls.|
The real ceiling fixture starts currentY30 and uses a triangle covering positiveZ.
The skew fixture instead starts previousY−30/currentY0 and registers the original
wall/ceiling triangles without claiming that the ceiling accepts. Its dynamic
edge→face sequence is the numerical diagnostic below.
The callback event string records virtual slot, polygon identity and global mode.
Dynamic callbacks use374/null. Static callback370/374 selects from the **retained
last polygon's normal**, and final314 follows callbacks. Every observation compares
contact, rootXYZW, face/edge/hitXYZW, last identity, callback/setter/pass/visit counts,
global mode, mode0/1 kernel counts and complete event order. Sequential queries use
the same synthetic actor; they do not prove registry order across real actors.
## First divergences and supported corrections
Text offsets are VA−400000; the source tools check actual PE mapping.
1. **Visited nodes and primitive tests have different owners.** At47B6A8..47B6B9
(file7B6A8),695D70 increments before4F69E5 is checked. The first managed
comparison reported0 visits versus original2 with geometry disabled.
`StaticCollisionContext.NodeVisits` now counts before that gate, while
`TriangleTests` continues counting only dispatched primitives. This is a
diagnostic-state correction, not a new collision response.
2. **Segment projection X retains its product.** The skew case's first differing
kernel is zero-based6: input(0,0,15,1) matches, but source hitX is
`8.68490133e-7` and managed hitX was0.4699A0..4699DE (file699A0) retains
directionX×projection until adding originX; Y/Z products have intermediate
float stores. `OriginalSphereResponse.NearestOnSegment` now follows those
separate boundaries.47C194..47C1D0 likewise retains X/Z products through the
final edge-root subtraction. This makes the complete step6 output match.
3. **Ray normalization and hit publication retain intermediates.** The next
mismatch is step7's face hit.469BE5..469C09 (file69BE5) uses OMT's returned
quick-magnitude quotient directly for reciprocal normalization; a float return
inserted an extra rounding.469DF0..469E13 (file69DF0) stores hit coordinates
after multiply+add. `OriginalBodySupport` now retains the magnitude intermediate
as double for these ray callers, and delays coordinate rounding until that
final store. Its public float QuickMagnitude result and caller-owned nearest
reduction remain float. All14 final observations and12 kernel steps now match.
Intermediate failure logs and experiment snapshots remain local, including
`managed-initial.log`, `managed-after-visits.log`, `projection-experiment.log`,
`ray-experiment.log` and `wide-experiment.log`. No epsilon or near-zero clamp was
added. The comparator parses native9-digit and managed shortest decimal spellings
back to float32; distinct values remain failures. Double intermediates reproduce
these fixtures, not every x87 extended-precision or table-boundary case.
## Implementation and validation
Runtime changes are confined to the shared support/ray and sphere kernels plus
the additional static visit counter. `tests/GamePort.DynamicReclipTests` keeps
fixed numeric references and all12 kernel steps for artifact-free normal runs;
`--native` checks the complete generated observation file, and `--dump` supports
future differential tracing. Python tests guard instruction boundaries, safe
generation and exact-float comparison semantics. Both probe source and expected
data are test fixtures, not new actor implementations.
`scripts/test.sh` exits0 with348 Python+2 HUD Python checks, all managed projects,
36 Vulkan/render checks with zero validation errors, audio lifecycle and539
original hashes. Debug build has zero warnings/errors. Native Linux SDL Carl
control (first neutral input/300 idle ticks) and VR04 jetpack steering/resize
also pass with zero validation errors. Windows x64 self-contained cross-publish
succeeds to a separate artifact directory. Existing packaged DLLs and BUILD-INFO
remain identical to924ad2f; no new package or Windows execution is claimed.
Reproduction after **waiting for disposable prefix preparation to complete**:
```sh
python3 tools/binary/dynamic_reclip.py ../Original --output artifacts/dynamic-reclip/native
i686-w64-mingw32-gcc -O0 -Wall -Wextra -Wno-unused-function -o artifacts/dynamic-reclip/native/probe.exe artifacts/dynamic-reclip/native/dynamic_reclip_probe.c
python3 tools/recon/reference_capture.py --asset-probe --probe-executable artifacts/dynamic-reclip/native/probe.exe --probe-success DYNAMIC_RECLIP_COMPLETE --seconds 45 --output artifacts/dynamic-reclip/probe-verified
python3 tools/binary/dynamic_reclip.py ../Original --output artifacts/dynamic-reclip/native --capture artifacts/dynamic-reclip/probe-verified
source scripts/env.sh
dotnet run --project tests/GamePort.DynamicReclipTests -- --native artifacts/dynamic-reclip/native/observations.json
```
For a complete field/step difference report, run `--dump` to a local file, then
pass that file as `dynamic_reclip.py --managed <file>` with the same output path.
The generated comparison explicitly records every mismatch; it does not turn
nearby values into parity. Final platform/suite receipts are in progress.md and
compatibility.md; local identities/results are in artifacts/dynamic-reclip/.
## Remaining gate and next chunk
Populated-grid composition now supports a bounded shared-context response. Before
live admission, recover actual Jimmy/VR04 actor callback mutations, live flag/
radius reads after callbacks, constructor-only helper re-entry/cleanup and real
physical registration order across LOAD/RETURN. Reentrant global-mode and grid
mutation, extreme coordinates/degenerate inputs and general x87 boundary behavior
remain open. Preserve lab/outside capsule and the existing VR04 static-only actor
set. Next focus on real actor callback ownership, then implement the bounded
response adapter against those interfaces. No movement tuning, angle reset,
gadget feature or unsupported dynamic actor is introduced here.