Record transparent geometry sort-driver one-shot

docs/object_matching_logs/rasterizer_transparent_geometry_obj_jonas_sort_driver_lifetime_20260825.md

@@ -0,0 +1,188 @@
+# `rasterizer_transparent_geometry.obj` sort-driver lifetime one-shot
+
+## Result
+
+This Jonas/Codex ledger records one bounded, failed-closed source-lifetime
+experiment for
+`source/rasterizer/rasterizer_transparent_geometry.obj::_code_00174120` from
+clean cumulative commit
+`b6d0ccd083ae7118ac57e57d8992eac89ef088c5`. The candidate made causal
+progress but did not satisfy strict semantic-COFF equality. It was removed
+immediately, the immutable baseline object was restored, and no implementation
+or matching credit survives.
+
+The unit remains 15/19 accepted functions. The residual remains 173 meaningful
+/ 176 padded bytes with 14 relocations.
+
+## Provenance and prior documentation
+
+The January split object is the byte and relocation authority. Its raw SHA-256
+is `d1bcf0d85a17c27f0cb9797931187fad486c2ac7b8ce7bcb43e706422c160722`.
+
+Before freezing the shot, the complete object-specific Claude research and the
+relevant compiler-allocation guidance were read from
+`claude/near-complete-objects-20260816` at
+`74a085e90ae66511c22820a8f5256b3273342542`:
+
+- `research/transparent_geometry_recovery.md`, blob
+  `52d38e05638dac26c4eb658a7849a4f5a68869b5`;
+- `research/transparent_geometry_sort_comparator.md`, blob
+  `1e5d6c93ed106ff9a7b9d694424884269d19e074`;
+- `research/register_tie_reachability.md`, blob
+  `1798df9e9e2fc5ade819a1d7b49d8fd84bc283e3`;
+- `research/register_tie_cumulative_findings.md`, blob
+  `0621651f58224429a724262580c9a7d25c2d42f4`.
+
+The first two records establish the sort-driver topology, the real in-TU
+private assert-helper relationship, the typed group and sorted-index storage,
+and the comparator call. The register documents establish that definition and
+lifetime order are legitimate source levers for integer allocation, while a
+synthetic probe cannot authenticate a whole-function schedule.
+
+The independent HCEA reconstruction at commit
+`0fd88176fe19df400576d2cbb447d3363080cb52` supplies the readable identity-fill,
+`qsort`, and rank-stamping topology in `src/rasterizer_sort_external.c`, blob
+`8b7fab90eb931cd6e35ed2874219fd28d7de75ab`, payload SHA-256
+`c186863a177fa5e40eabbd9860d6f9d4af54286942b0359f6e5ff81d7576c056`.
+HCEA is a later Xbox 360/PPC reconstruction and is not a January-PC byte
+oracle; it supports only the high-level source topology.
+
+The current production object was reconstructed after Claude's target-only
+frontier measurement. Reachable history and workspace census contained no
+recorded compile of this exact direct-expression, block-local-pointer, and
+pre-test-index lifetime combination.
+
+## Frozen baseline
+
+The clean source identity was Git blob
+`f278c3493a9aeb17d83e3d5cdd5aa382183459f4`, with CRLF working-tree SHA-256
+`874d31a32c82f5ac75330c668444e44cd4a12fd5fe12f84305a7e4badb6bfd5a`.
+The isolated worktree was configured with the repository's XDK 3911 graph and
+Microsoft 32-bit C/C++ Optimizing Compiler 13.00.9254.1. `CL.Exe` SHA-256 is
+`483e00c47bb08d699475a642bcff15b5b2036350b31c540e88a506baf101da11`;
+the production flags remained `/O2 /Oy- /DDEBUG /Dxbox` and the generated
+include closure.
+
+A clean pre-edit baseline compile established the local Ninja command record.
+The frozen baseline object and immutable copy both have raw SHA-256
+`906223ee9cf34ea1186f292d3ca1d4c92a031ea54918ca3a419bc731ce303581`.
+The regression manifest
+`build/audit/rasterizer_transparent_sort_driver_baseline_20260825.json` is
+400,979 bytes with SHA-256
+`7997a12c7fd4139c7f769612dcfc04a92c15ac60b879a636d99d27ce17c354aa`.
+Its self-check passed with no failures or warnings and froze these 15 inherited
+accepted functions:
+
+- `_code_00173f70` and `_code_00173fa0`;
+- `_rasterizer_transparent_geometry_begin`, `_dispose`, and `_end`;
+- `_rasterizer_transparent_geometry_get_group_from_presorted_index`;
+- `_rasterizer_transparent_geometry_get_group_pending_status`;
+- `_rasterizer_transparent_geometry_get_group_presorted_index`;
+- `_rasterizer_transparent_geometry_get_groups` and `_get_groups2`;
+- `_rasterizer_transparent_geometry_get_primary_vertex_type`;
+- `_rasterizer_transparent_geometry_new_group` and `_new_group2`;
+- `_rasterizer_transparent_geometry_set_group_pending_status` and `_stop`.
+
+The residual's target and baseline fingerprints were:
+
+| Side | Meaningful / padded | Relocations | Normalized SHA-256 |
+| --- | ---: | ---: | --- |
+| January target | `173 / 176` | 14 | `4f78b8476535a901325781b67d7ed04665e59478fb419ddd0c607c95220655d5` |
+| clean baseline | `173 / 176` | 14 | `30cd227ef42506ffe5a4050491eff57faef6a200cbffc05301077bca69f40a50` |
+
+Their normalized bytes differed at 69 offsets, first at `+0x05`. The target
+saved only ESI/EDI at entry, preserved the first count in EAX, zeroed the second
+pass index before testing the count, and shrink-wrapped EBX around the
+nonempty second pass. The baseline saved EBX at entry, clobbered the first EAX
+count with the named group pointer, initialized the second index after the
+test, and loaded the sorted-index pointer before the group pointer.
+
+## Frozen candidate and one production compile
+
+The one candidate made three ordinary-C lifetime changes in the existing body:
+
+1. removed the function-scope `group` pointer;
+2. passed `transparent_geometry_groups + group_index` directly to the private
+   assert helper in the first pass;
+3. initialized `group_index` before the post-sort positive-count test and
+   declared `group` before `group_sorted_indices` inside that positive block.
+
+The candidate source was Git blob
+`53da9b0b8fd6613f0d41c34c4ac0b6a8d4186e6c`, with working-tree SHA-256
+`e317b4b434495c7fc6515bdd5817b49f9dfd34a69d2bac1c34e1f336f4311f15`.
+It used only named typed members and ordinary control flow. It introduced no
+assembly, `volatile`, `register`, pragma, annotation, raw address, byte offset,
+pointer/integer reconstruction, object-representation cast, type pun, undefined
+behavior, synthetic anchor, or compiler option change. The file's inherited
+packing pragmas were unchanged and outside the patch.
+
+One XDK `/Zs` invocation with the exact generated include closure passed and
+left both baseline object hashes unchanged. The following dry run exposed
+exactly one code-producing edge:
+
+```text
+[1/1] CL build\base\source\rasterizer\rasterizer_transparent_geometry.obj
+```
+
+Exactly that edge received the wave's sole production candidate compile. The
+immutable first-shot object has raw SHA-256
+`5f8e8cf790033dc6af3e78b64c253b59d676c2e47ee741537fa5f5a43612d51a`.
+
+## Measured rejection
+
+The candidate retained the target's `173 / 176` size and 14-relocation count,
+but its normalized SHA-256 was
+`c7a86bc76b0c47bf55982cca4a409ba259335ac1840eecf088b53392dc425446`,
+not the target hash. The normalized distance improved from 69 to 58 differing
+bytes and the first divergence moved from `+0x05` to `+0x11`, but strict
+equality is binary and no fuzzy credit is claimed.
+
+The change did explain part of the target:
+
+- the entry prologue now saved only ESI/EDI;
+- the second-pass index was zeroed before the count test;
+- EBX was shrink-wrapped around the nonempty second pass;
+- the loop body from `+0x90` through the epilogue aligned with January.
+
+It did not reproduce the remaining schedule. In the first pass January formed
+the transient group in ECX from a base held in EDX, while the candidate formed
+it in EAX from a base held in ECX. Around `qsort`, January pushed the comparator
+and width before loading the sorted-index pointer; the candidate loaded that
+pointer first. The candidate's post-sort prefix was therefore one byte later
+until alignment at `+0x90`.
+
+Relocation addresses make the residual explicit:
+
+| Side | Relocation addresses |
+| --- | --- |
+| target | `1,18,40,45,50,57,62,71,88,96,102,108,123,130` |
+| baseline | `1,18,40,45,50,57,66,75,89,94,103,109,122,128` |
+| candidate | `1,18,40,45,50,57,66,75,89,94,103,109,124,131` |
+
+The full regression check reported every inherited accepted function as
+`still_exact` and no newly exact function. The only non-code change was the
+compiler-owned anonymous `.debug$F` section caused by the source-local change;
+no runtime-data or symbol-ownership gain is claimed. Because the intended
+function remained nonexact, no debug adjudication, full Halo/libcmt validation,
+status change, or replay chain was permitted.
+
+## Restoration and reopen boundary
+
+The candidate was inverse-reverted without a source-shape retry. `apply_patch`
+changed the working-tree line-ending representation, so the clean cumulative
+copy was used solely to restore the CRLF bytes. The live source is again Git
+blob `f278c3493a9aeb17d83e3d5cdd5aa382183459f4` and raw SHA-256
+`874d31a32c82f5ac75330c668444e44cd4a12fd5fe12f84305a7e4badb6bfd5a`.
+The immutable baseline object was copied back; live and immutable raw hashes
+again equal
+`906223ee9cf34ea1186f292d3ca1d4c92a031ea54918ca3a419bc731ce303581`.
+A one-target Ninja dry run reports no work, the manifest check again passes
+with all 15 inherited functions `still_exact`, and the tracked tree contains
+only this ledger.
+
+Do not retry the compiled direct-expression + block-local-pointer + pre-test
+index family, or split it into spelling variants, without genuinely new
+January source or another compiler-authentic artifact that explains both the
+first-pass ECX/EDX assignment and the pre-`qsort` relocation order. HCEA's
+semantic loop topology and allocation reachability alone are now exhausted for
+this source family.