Cesarus85--mjolnir-vr/docs/object_matching_logs/cluster_partitions_obj.md

source/structures/cluster_partitions.obj strategy ledger

This retroactive ledger preserves the evidence available from Git, config/parked.json, source comments, and fresh strict COFF measurements. The original lane predates mandatory experiment matrices; missing per-variant hashes are called out rather than reconstructed.

Scope and baseline

• Compiler: XDK 3911 CL.Exe 13.00.9254.1, /O2 /Oy- /DDEBUG /Dxbox.
• January target: build/split/source/structures/cluster_partitions.obj.
• Current inventory: 14/14 target functions strict-exact; 1,808/1,808 padded code bytes, with every relocation address, type, destination, and addend exact.
• All 14 target-owned .rdata COMDATs are strict-exact: 476/476 logical bytes (500 aligned bytes in the ordinary report), zero relocations. The target object owns no .data or .bss.
• _reference_list_copy was the sole residual and is now resolved.

Historical parked residual (resolved)

Function	Size T/B	Relocs T/B	Hash T/B	Measured residual	Class
_reference_list_copy	0xC0/0xC0	9/9, including addresses	28ab903006bf625bb291ffe9b82a710eabf80400b9d1f2403b78b04429904c8b / historical dcebabeded0e38211be7325fc38d98591a621f800556829f6184e38dbacf7c03	VC7 assigned source and destination pointer live ranges to opposite registers.	resolved by coalescing order

Preserved experiment history

Eight legal-C source shapes were measured by the historical closeout lane. The surviving record proves they retained equal size and relocation identity without breaking the pointer-register mirror. The exact eight source diffs and transient hashes were not preserved, so this ledger deliberately does not invent them.

Do not repeat blind declaration-order or source/destination pointer-alias spelling sweeps. Do not use volatile, assembly, byte patches, undefined behavior, or compiler-flag changes.

Resolution and provenance

Claude-assisted commit d84c4d13 established the reachable legal-C control. Both source-order changes are required:

• initialize absolute_index before deriving the cursor pointers, so the counter claims ESI before the source value dies; and
• advance source_reference before result_reference, preserving the source cursor in EBX while the result cursor spills.

That produces the target hash 28ab903006bf625bb291ffe9b82a710eabf80400b9d1f2403b78b04429904c8b with 192/192 padded bytes and 9/9 exact relocations.

Full-object admission audit (2026-08-20)

Fresh isolated XDK 3911 rebuild and the hardened regression fingerprint:

• target functions: 14; candidate shared functions: 14; strict-exact: 14/14;
• padded target code: 1,808/1,808 bytes;
• target-owned non-code: 14/14 .rdata sections, 476/476 logical bytes, exact normalized hashes, symbols, flags, and relocation identity;
• target-owned .data / .bss: absent.

The candidate also emits three unreferenced out-of-line inline-helper COMDATs (reference_list_new, reference_list_get_next_datum_index, and reference_list_add) and four pooled string COMDATs. These are not unique object-owned content:

• every extra candidate section uses IMAGE_COMDAT_SELECT_ANY;
• no relocation in any of the 14 target functions calls the three extra helper bodies because all calls were inlined;
• the selected January owners are source/hs/object_lists.obj for the three helpers and two helper strings, source/objects/objects.obj for "location", and source/camera/editor_flying_camera.obj for "position";
• every candidate duplicate is freshly strict-exact against that selected January owner, including helper relocations.

This is the documented discardable-COMDAT case, not extra ordinary code or data. Parameter declarations follow the house rule (one parameter per line); the exact function source retains explicit final returns and the target-required early exit in reference_list_remove. Typed tag/object access rules do not apply because this translation unit performs neither kind of raw access.

Validation completed before admission

• Forced cluster_partitions.obj XDK compile: exit 0.
• Hardened function comparator: all_equal: true, 14/14 functions.
• Fresh target-owned section/ownership fingerprint: all 14 runtime sections strict-exact, with no target-only section.
• Full halobetacache_build: exit 0 after 466 actions.
• Full libcmt_build: exit 0 after 102 actions.
• Strict semantic report: 470 units scanned, 3,604 functions evaluated, zero unit errors; the admission audit found cluster_partitions as the sole finished-but-unmarked candidate with zero function and data gaps.

Post-admission gates

• ninja progress: exit 0. Halo progress became 3,349/7,574 accepted functions, 406,188/1,770,166 code bytes, 1,792,466/3,923,451 data bytes, and 268/468 objects.
• Post-admission audit: no finished-but-unmarked candidate remains and no cluster_partitions contradiction was reported. The unrelated pre-existing shell_xbox contradiction remains outside this lane.
• Parked-evidence validation: 6 active, 0 stale, 0 invalid.
• Fail-closed repeat-rebuild regression fingerprint: target evidence, accepted-function fingerprints, every non-code section fingerprint, symbol ownership, and comparison states were unchanged. The complete-object SHA changed only because MSVC stamps the COFF header; the regression gate does not treat that non-section timestamp as acceptance evidence.
• Comparator, regression, semantic, admission, and parked-evidence tests: 98/98 pass. Full tools test discovery: 179/179 pass.
• Diff-policy audit: exactly five object-scoped tracked files changed; the only Markdown change is this ledger; no frozen object or Claude Markdown path is changed or deleted; git diff --check passes.

Disposition: strict-complete, full-build clean, regression-stable, and admitted as Matching.