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

players.obj strict-closeout strategy log

Baseline

• Lane: codex/players-closeout-root at caa820ccd (0892e393b, 86a73bc55, caa820ccd).
• Toolchain and flags: repository XDK 3911 CL, unchanged /O2 /Oy- /DDEBUG /Dxbox.
• Strict comparator: tools/coff_compare.section_infos_equal against build/split/source/game/players.obj.
• Current result: 106/110 sections exact and 16,924/21,164 padded section bytes exact.
  • Functions: 66/70, 11,664/15,904 padded bytes.
  • Data: 40/40, 5,260/5,260 padded bytes.
• config/config.json remains NonMatching.

Residual	Target size/relocs	Target normalized SHA-256	Current normalized SHA-256	First meaningful divergence
_code_000a9ff0	0x170 / 22	9e3ab79a5087c8d11b97f52c30afc3f596faa4aa6122480dc9f16d247be8ca9f	d9ca587214fd10308b7a16a797d8a154102f2095fbf28324cb9e3e2da8d8483e	+0x2E: January forms &iterator before homing the activation result; current CL homes EAX first.
_code_000aa9e0	0x510 / 61	8bb48b429649a3afcd41cf238ca9b1d5dba0ff61e15f2205f72512918cad7c43	740c065b9b622d07827e6ae83ac2d79aeb3813a54bbf7b8a43c7c0c8f0e8b36f	+0x5: January uses a 0x7C frame; current CL uses 0x78, then chooses a different local/register topology.
_players_update_before_game	0x6E0 / 101	8c92baaaebf7177d7bc54239c21b0923703a50f3ede19dca3a81a457f2860947	51c2c7b42d12188a024fac2a2c5d030c194d6a91be0b4e3d57bdd5efe1dd1c8c	+0x62F: four register-encoding bytes; January uses EDX for the unit index and EAX for the next iterator address.
_players_update_after_game	0x330 / 52	af64f40efdb1be0cd2754ada85b34acecdd4480a39c73b5a9ef2938c2c1d315b	6c0d80f1ab3c18aa8447682e0e9e7365486c42df02072c1a24c729a54306d7b7	+0x74 branch displacement; the causal difference begins at +0x18C, where January assigns EDI to the BSP loop index and EBX to the element pointer, opposite current CL.

Bounded source experiments

All experiments retained the ABI and flags and were rebuilt as the single build/base/source/game/players.obj target. Unless marked accepted, the source was reverted after measurement.

Function	Evidence-based source shape	Strict outcome	Status
_code_000a9ff0	Fold activation-result assignment into data_iterator_new's second argument with a comma expression.	0x170 / 22; canonicalized to current hash d9ca5872...; same +0x2E scheduling window.	Reverted
_code_000aa9e0	Split the unit-index/biped aggregate into independent typed locals.	0x510 / 61; hash 0d30bbd7f96442016f38bdfc9c200999b37e61291f2c7d401786153f876f2b3b; frame remained 0x78, unit and biped homes moved to -0x10/-0x14, not January's -0x14/-0x10.	Reverted
_code_000aa9e0	Extend the placement aggregate with the source-root pointer to model January's three adjacent homes.	0x510 / 61; hash e79a85eafa5ab7123747724dc7b241a69307a6631c7a103edc09e4532161389a; recovered the 0x7C frame but placed the aggregate at -0x30, first differing home at +0x24.	Reverted
_code_000aa9e0	Wrap only the source-root pointer in a one-field struct, then in a one-element typed array.	Both scalarized to current 0x510 / 61 hash 740c065b...; no stack home or code change.	Reverted
_code_000aa9e0	Move the three-field placement aggregate after the matrix/vector locals.	Canonicalized to the same 0x510 / 61 hash e79a85e...; declaration position did not move the -0x30 aggregate home.	Reverted
_code_000aa9e0	Group the source-root and later source-biped pointers, then take a non-volatile typed address of the mutable source-unit parameter.	Both scalarized to current 0x510 / 61 hash 740c065b...; neither forced January's -0xC home.	Reverted
_code_000aa9e0	Materialize the initial biped_get result through a typed temporary before copying it to the placement state.	Copy-before-result canonicalized to 740c065b...; result-before-copy moved xor bl,bl ahead of the call and produced hash 11d6457990632220105232975d079705b646a90e8f14c0bcb8369377236a0fa8.	Reverted
_players_update_before_game	Move _ReadWriteBarrier() before the local unit-index assignment.	0x6E0 / 101; hash af6add6c8b40f7ae93e5853584e2f075ce3a52c88707cddefbe94c5a6543cbf4; regressed from four to ten normalized bytes, first at +0x62B.	Reverted
_players_update_before_game	Make the local unit index const.	Canonicalized to current 0x6E0 / 101 hash 51c2c7b4...; same four bytes.	Reverted
_players_update_after_game	Move the loop index and element-pointer declarations across inner/function scopes and reorder declarations.	Every form canonicalized to current 0x330 / 52 hash 6c0d80f1...; same EBX/EDI mirror.	Reverted
_players_update_after_game	Remove the barrier after global_scenario_get.	0x330 / 52; hash 20ac0c082a00bbe619f3f3c868fbc6c6933177c5ba04d1ad46ae3613be37f18d; regressed to the base-copy/add scenario-block expression and 373 normalized bytes.	Reverted
_players_update_after_game	Move that barrier after the scenario block-address assignment.	0x330 / 52; hash 0f8682e9333c2a6c9cdf18fae3a4a77cc7d36868cd0badb9069bec177c635c2b; retained a non-target copy/add sequence beginning at +0x184.	Reverted

No source experiment improved a strict section in this closeout pass. The production source is restored to caa820ccd.

Park decisions and next leverage

• _code_000a9ff0 is parked as methodology class D (scheduler/lifetime). Twenty-one of 22 relocation addresses are exact; only player_data moves three bytes inside the +0x2E..+0x39 scheduling window, and all following instructions reconverge. Reopen only with provenance for the activation result/iterator statement shape.
• _players_update_before_game is parked as class C (register fixed point). All 101 relocation identities and addresses are exact. The only differences are four ModRM/register bytes at the final unit_control and next-iterator sequence. Reopen only from a caller/compiler forced-register anchor.
• _players_update_after_game is parked as class C/D. All 52 relocation identities agree; 33 addresses are exact and the remaining addresses form two uniform drift plateaus (+1 for six, then -5 for thirteen) after the EBX/EDI loop mirror. Reopen only with source provenance that changes the loop index/element live-range weights while retaining the required post-scenario barrier shape.
• _code_000aa9e0 is parked as an unresolved class A/C boundary, not claimed as a proven compiler-only tie. All 61 relocation identities and the 0x510 padded size agree, but only 20 relocation addresses are exact and the target's opening 0x7C frame/local homes remain unrecovered. January homes player, player_unit_index, biped, and source_root_object at -0x8, -0x14, -0x10, and -0xC; current CL uses a 0x78 frame and keeps the source-root pointer in ESI. Reopen only with new donor/PDB/local-variable provenance or a legal-C source form that recovers all four homes; more declaration/aggregate permutations are not evidence-based.
• A strict completion still requires all four current hashes to equal their target hashes plus a fresh 110-section sweep. Equal size/relocation count is not match credit.

Do not repeat

• _code_000a9ff0: explicit iterator pointer, result temporary, result/iterator aggregate, comma-expression assignment, or compiler barrier around the activation call. These retain the same store/load/LEA schedule or regress it.
• _players_update_before_game: direct player->unit_index, barrier removal or relocation, const/register, pointer aliases, and declaration-order changes. The four-byte baseline is best.
• _players_update_after_game: barrier removal/relocation, declaration-scope permutations, register hints, and equivalent for declaration placement. Keep the barrier immediately after global_scenario_get.
• _code_000aa9e0: independent unit/biped locals, a three-field placement aggregate in either declaration region, one-/two-field source-pointer wrappers, a non-volatile source-parameter alias, an initial biped temporary, declaration reordering, and volatile pointer/parameter access. Volatile forms are also barred by current methodology.
• Do not change optimization flags, add inline assembly/object patches, accept fuzzy objdiff percentages, or mark the TU Matching before the hardened 110-section gate passes.

2026-07-29 isolated follow-up

The collision-safe follow-up is preserved on codex/players-final-four as source checkpoint 576d93f33 and full ledger commit 8dad10254. The source checkpoint is deliberately not integrated here because players.obj remains incomplete; these measurements prevent later lanes from repeating the same work.

• P01: rewriting the final iterator loop from while to explicit for produced no bytes and was reverted.
• P02: a Players-only long declaration for objects_get_activating_cluster_index modeled January's measured cross-TU return-ABI bug and reduced _code_000a9ff0 from 287 to 11 normalized differing bytes. The lane labels the preserved bug and documents the corrected-build fix.
• P03: moving the activation-result declaration before the iterator had no effect and was reverted.
• P04: the HCEA-PDB-backed player_teleport_internal local topology retained 0x510/61 and reduced _code_000aa9e0 to 929 normalized differing bytes.
• P05: restoring the scenario barrier, two BSP counters, and measured loop topology retained 0x330/52 and reduced _players_update_after_game to 222 normalized differing bytes.
• P06/P07: the historical 877-byte teleport candidate was reproduced, but it depended on an unproven volatile pointer-store expression used only as a codegen lever. It was rejected under campaign policy; the clean PDB-backed P04 form was restored.
• P08: recovered field/flag names and compile-time January layout assertions were byte-neutral across the whole TU.

Final isolated-lane gate: 106/110 sections, 66/70 functions, and 40/40 data sections exact. The four residuals retain equal padded size and relocation count but receive no exact credit:

Function	Size / relocs	Candidate SHA-256	Differing bytes
_code_000a9ff0	0x170 / 22	d9ca587214fd10308b7a16a797d8a154102f2095fbf28324cb9e3e2da8d8483e	11
_code_000aa9e0	0x510 / 61	d833e6d8f06cfdb556417d852c6a0717eecfdb2726f11a92d9241c15b363bbb0	929
_players_update_after_game	0x330 / 52	6c0d80f1ab3c18aa8447682e0e9e7365486c42df02072c1a24c729a54306d7b7	222
_players_update_before_game	0x6E0 / 101	51c2c7b42d12188a024fac2a2c5d030c194d6a91be0b4e3d57bdd5efe1dd1c8c	4

P20 compiler-provenance and patch-level falsification

The January cachebeta.pdb module stream for \halo\objects\halobetacache\players.obj contains a 60-byte S_COMPILE2_ST record. Its front-end and back-end version triples are both 13.0.9254, followed by the producer string Microsoft (R) Optimizing Compiler. This is target-authored evidence that the TU was compiled with the same numbered VC7 build used by the campaign.

The unchanged retained source was then compiled once under the only two distinct local VC7 compiler binaries:

Compiler	_code_000a9ff0	_code_000aa9e0	_players_update_before_game	_players_update_after_game
13.00.9254.1	d9ca5872...	d833e6d8...	51c2c7b4...	6c0d80f1...
13.00.9210	d9ca5872...	d833e6d8...	51c2c7b4...	6c0d80f1...

All four normalized candidate hashes were byte-identical across compiler binaries. The older 9210 compiler therefore supplies no code-generation control for this TU, while the target record materially weakens the hypothesis that an unidentified compiler build explains the residuals.

Disposition: no source retained and no exact credit. Do not repeat a local compiler census or download nominally different 3911 kits without first proving that their CL.Exe hash is new. The remaining reopen conditions are authentic January/cross-build statement topology or a focused legal-C control that changes the measured register/lifetime window without moving any exact relocation.