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

breakable_surfaces.obj exact-match strategy log

Scope and provenance

• Source TU: source/physics/breakable_surfaces.c
• Target object: source/physics/breakable_surfaces.obj
• Target build: Halo January 2002 Xbox build
• Compiler: XDK 3911 Microsoft C/C++ 13.00.9254.1
• Flags: /O2 /Oy- /Ow /QIfist /DDEBUG /Dxbox
• Comparator: hardened COFF function and section comparison
• Status: rigorously fuzzy-parked at the current natural-source frontier; no strict matching credit has been granted

The rough October source at C:\Users\isabe\Downloads\halo-main (1)\halo-main\src\halo\physics\breakable_surfaces.c is a topology hint only. January machine code, relocations, and data ownership remain authoritative.

Current validated state

• Functions: 11/12 strict exact.
• Only residual: _breakable_surface_effect.
• Target/current padded size: 0xFC0/0xFC0 (4032/4032).
• Target/current relocations: 117/117, with the same semantic destinations.
• Target/current frame: 0x1240/0x1240.
• Target normalized SHA-256: 510486a7d0a1f2fc1dfb8f0c47990c57f83ce1030ef5e4db8dfc42943dc56bfa
• Current normalized SHA-256: 678a8125d160643a9f1226240128c662bcbb85e7f968f8fd3ed0f0b411f77797
• All 11 protected sibling functions remain strict exact.

The current source is the 2026-08-31 Claude reconstruction reconciled against the current shared headers and September house rules. It supersedes the E09 traversal/audio scratch union checkpoint described below. The older experiment ledger remains here because it is still useful negative evidence.

Historical checkpoint controls (superseded)

E09 and E18 were legitimate intermediate evidence, but they are no longer the production source shape. The final reconciliation recovered the later natural block lifetimes and the same rectangle2d field meanings without retaining the scratch union.

E09: traversal/audio scratch union

January reuses one 0x58-byte region for traversal scratch and the later sound_location. The retained C89 union models the measured layout:

• traversal s_offset, t_offset, two projected points, and eight 2D vertices;
• audio view with sound_location at union offset 0x2C;
• both views end at the same byte.

This is defined lifetime-based storage reuse, not aliasing or byte padding.

E18: rectangle2d grid bounds

Replacing four unrelated short locals with rectangle2d grid_bounds moved the candidate from 0xF90 to the exact target size 0xFC0 while preserving 117/117 relocations and every exact sibling.

The native structure layout is exactly the order proved by January:

Structure field	January meaning	January home
grid_bounds.y0	t_min	-0x5C
grid_bounds.x0	s_min	-0x5A
grid_bounds.y1	t_max	-0x58
grid_bounds.x1	s_max	-0x56

This is an evidence-backed type recovery and remains in production even though the function is not yet exact.

Current stack/codegen finding

E18 makes every large aggregate start exactly 16 bytes above the January home:

Region	Target	E18
surface_queue[1024]	-0x1240	-0x1230
surface_vertices3d[8]	-0x240	-0x230
traversal/audio scratch	-0x1E0	-0x1D0
new_particle_data	-0x188	-0x178

The missing 16 bytes are four January dword integer-conversion homes. For each grid bound January emits:

call  _ceil/_floor
fstp  dword ptr [ebp-real_home]
add   esp, 8
fld   dword ptr [ebp-real_home]
fistp dword ptr [ebp-distinct_integer_home]
mov   r16, word ptr [ebp-distinct_integer_home]
mov   word ptr [ebp-grid_bound], r16

E18 instead emits fistp qword to one shared compiler temporary. The target therefore rounds through single precision and owns four distinct dword conversion homes; the present compiler forwards through every normal readable-C spelling tested.

Experiment matrix

Every reverted experiment preserved all 11 exact siblings unless explicitly noted.

ID	Source or compiler experiment	Result	Decision
E01	Move position and position_2d_test inward	byte-neutral	reverted
E02	Move both vertex arrays into the while scope	byte-neutral	reverted
E03	Declare the 2D vertex array later	byte-neutral	reverted
E04	Remove the duplicate queue-read shadow	byte-neutral	reverted
E05	Hoist random offsets to function scope	byte-neutral	reverted
E06	One named real between ceil/floor and cast	byte-neutral; store forwarded	reverted
E07	Group grid bounds and loop indices	byte-neutral	reverted
E08	True sibling traversal/audio scopes	frame 0x1228; proved lifetime reuse	superseded by E09
E09	Meaningful 0x58 C89 scratch union	size 0xF90, frame 0x122C	retained
E10	Remove explicit long from the grid casts	byte-neutral	reverted
E11	Static grid-index helper	caller byte-neutral; unwanted helper COMDAT	reverted
E12	Named real plus four named long homes	optimizer eliminated homes	reverted
E13	Aggregate containing rounding/conversion fields	optimizer eliminated homes	reverted
E14	Address escape / compiler-barrier shape	did not produce target chain	reverted
E15	Conversion-home array plus barrier	did not produce target chain	reverted
E16	Pointer helper plus barrier	did not produce target chain	reverted
E17	__forceinline float ceil/floor wrappers	caller byte-neutral; two unwanted COMDATs	reverted
E18	Native rectangle2d grid_bounds	exact 0xFC0, 117/117; correct bound layout	retained
E19	real_rectangle2d staging plus E18	byte-identical to E18	reverted
E20	Four named long homes plus E18	byte-identical to E18	reverted
E21	Historical function-scope bounds plus long loop indices	regressed to 0xF70; hash f32f3d8f...	reverted
E22	Move only rectangle2d grid_bounds to function scope	byte-identical to E18	reverted
E23	Change only the two grid loop indices to long	same 0xF70 regression as E21	reverted
A01	Compile E18 as C++ (/TP)	function byte-neutral; decorated symbols	analysis only
A02	/TP with explicit ceilf/floorf	unwanted helpers; caller byte-neutral	analysis only
A03	Remove /Ow	byte-neutral	analysis only
A04	Enable /Op	4320 bytes, 114 relocs; broad regression	rejected
A05	Minimal direct/named/four-value float-to-short fixtures	every form collapsed to shared fistp qword; no target motif	analysis only
A06	Four legal union-backed long/short conversion objects	also collapsed to shared fistp qword	analysis only

Corpus and compiler audits

• A relocation-aware scan of all 833 January split objects found the complete float-home conversion motif only eight times: four here and four in nonexact _build_structure_lens_flares.
• A scan of 467 rebuilt base objects found zero occurrences.
• _build_structure_lens_flares is not a reconstructed-C donor: target/base is 4336/4000 bytes and 156/157 relocations.
• All locally available compatible compiler sets are byte-identical VC7 13.00.9254.1 binaries. No alternate compatible compiler/QFE is locally available.
• Upstream history commit 95e6477328dd7742175074de5eab0a20c3e7bde0 provided a real source-shape lead: function-scope bounds and long grid indices. E21-E23 showed that bound scope is neutral and the historical long index type causes the same 0xF70 regression, so it is not the January topology.
• The reproducible round_probe.c fixture proves that direct, named-real, named-long, four-output, and four-longs-first forms all lower through one shared fistp qword home under the campaign compiler.

Reproducible evidence is under research/breakable_surfaces_closeout/, especially STACK_SLOT_LIFETIME_MAP.md and donor_census/REPORT.md.

Do not repeat

• E01-E17 and E19-E23, including cosmetic variations with equivalent lifetimes.
• C++ mode, ceilf wrappers, /Ow removal, or /Op.
• Adding guessed padding merely to force the missing 16 frame bytes.
• Treating _build_structure_lens_flares as an exact donor.
• Inline assembly, volatile byte forcing, undefined behavior, object-byte patches, or compiler-flag changes.

2026-07-28 provenance audit: January PDB, HCEX PDB, and local source corpus

A second provenance pass found a strong source-semantics lead but no new legal-C code-generation lever.

January PDB negative result

C:\Users\isabe\Downloads\Halo Combat Evolved (Jan 14, 2002 prototype)\cachebeta.pdb names \halo\objects\halobetacache\breakable_surfaces.obj as module stream 154. The complete module stream is only 132 bytes: one S_OBJNAME record, one compiler record (Microsoft (R) Optimizing Compiler), and four zero tail bytes. It contains no function, local, type, or line records, so it cannot recover the grid-bound declarations, any rounded real, or the four conversion homes. The raw compiler record must not be treated as proof of an unrecorded precision flag.

HCEX source-line and code-generation evidence

The later debug and release symbol files are:

• C:\Users\isabe\Documents\Codex\2026-07-13\i-w\research\hcea_jun2011_prototype\payload\Halo CE Anniversary (Jun 24 2011)\HCEX.pdb
• C:\Users\isabe\Documents\Codex\2026-07-13\i-w\research\hcea_jun2011_prototype\payload\Halo CE Anniversary (Jun 24 2011)\HCEX_Release.pdb

They identify breakable_surface_effect as source d:\projects\code\hcex\sources\physics\breakable_surfaces.c, source MD5 CEACA3EE6F74879855CD2005D8B6C1AC. Debug RVA/length/frame are 0x00D0C1A0/0xE68/0x1360; release RVA is 0x01743190 with the same length/frame. Neither PDB contains injected source.

DIA line records map the four bounds to four consecutive source statements: lines 423-426 are the two ceil assignments followed by the two floor assignments; line 421 is the nonzero-density test and line 440 begins the grid loop. There are no intervening source lines for a separate named-real assignment/conversion pair. The surviving named-local records also contain no rounded-real/grid aggregate or scalar grid-bound names, although optimized-away locals cannot be ruled out from that absence alone.

PowerPC disassembly of the decoded debug image at flat RVA 0xD0C924 is stronger semantic evidence. Each libm result is immediately rounded from double to float with frsp (first at 0xD0C954), then converted with fctiwz/stfd and sign-extended to a short; the four results remain scalar GPR values. This is consistent with four direct (short)(real)ceil/floor(...) assignments and inconsistent with a required separate named real statement. PowerPC cannot distinguish the optional intermediate (long) because both int and long are 32-bit there.

Source-corpus and macro negatives

A hash census of 747 local breakable_surfaces.c files found 11 unique contents and no copy matching HCEX MD5 CEACA3EE6F74879855CD2005D8B6C1AC. Repository history contains reconstruction commits only, not an authentic Bungie source blob. The October file at C:\Users\isabe\Downloads\halo-main (1)\halo-main\src\halo\physics\breakable_surfaces.c is a decompiler reconstruction: it defines replacement breakable_floor and breakable_ceil functions and has grid comments/order inconsistent with the machine-code topology, so its named float val is not provenance.

The active type and macro audit found typedef float real in source/math/real_math.h and ordinary ternary PIN in source/cseries/cseries.h; no real-to-short rounding macro exists in the recovered source. The closest source sibling, source/structures/structure_lens_flares.c, spells all four bounds directly as (short)(real)ceil/floor(PIN(...)), and its January target has the same float-home conversion motif, but that function is itself nonexact and cannot serve as a strict donor.

Verdict and do-not-repeat expansion

The highest-likelihood original spelling is four direct scalar assignments with an explicit (real) cast, probably without the reconstruction-only (long). That spelling, named-real forms, named-long homes, aggregates/arrays, address escapes, helpers/wrappers, scope changes, C++ mode, and precision-flag probes have already been measured and do not make the campaign VC7 emit January's chain. No genuinely untried legal-C hypothesis now satisfies the reopen criterion. Do not repeat the January-PDB private-record search, HCEX injected-source/source-hash search, October-source audit, macro/type audit, or local source-hash census; reopen only if an authentic source blob, a distinct compatible compiler/QFE, or a strict donor artifact appears.

Residual classification

Compiler-precision/provenance fixed point. Size, call set, relocation count, recovered grid type, and large aggregate layout are strongly supported. January has a repeated single-precision rounding and four-home conversion topology that the current VC7 binary does not emit from any tested legal-C shape. This is not strict exactness and receives no completion credit.

Reopen/continuation criteria

Reopen only with evidence for one of:

1. an original October/January/HCEX source or debug topology that explains the intermediate real materialization;
2. a strict reconstructed-C donor for the complete conversion motif;
3. a genuinely new legal-C dataflow that predicts both the fstp/fld single-precision round trip and four distinct dword homes.

Disposition

NonMatching / rigorously parked at 11/12 exact functions. E09 and E18 were evidence-backed improvement checkpoints, not object completion. The newer reconciliation below is authoritative. Do not mark the object complete or grant credit from equal size or semantic plausibility.

2026-09-01 Claude-frontier and house-rule reconciliation

Corpus audit

The closeout reviewed all 1,648 registered worktrees and reduced their breakable_surfaces.c files to twelve distinct source variants. The unique complete frontier was the history on claude/breakable-surfaces-20260830, in particular commits 7fbe2a94d, a324c519b, 39207d9ec, a5af3125d, ba1109b49, and 230dc0af0. It was checked against the separate exact-breakable-surfaces-closeout lane, the canonical/object log history, research/breakable_surfaces_closeout/, the January target, the HCEX PDB and PowerPC implementation, the HCEA donor, the October reconstruction, and the local same-compiler donor census. No worktree contained a later credible frontier.

The retained source recovers the complete surface traversal and projection, particle-grid generation, outward/forward impulse shaping, particle tint and alpha evaluation, and optional break sound. It also retains the evidence-backed natural block scopes, rectangle2d grid bounds, real_point3d vertex lvalues, semantic private helper names, grouped j/k plane-distance association, and the target alpha multiplication order. The old duplicate TU-local conversion helper was removed: all four conversions now use the authenticated shared fast_ftol in source/cseries/cseries.h. That two-instruction x87 helper is independently corroborated by the HCEX PDB and January's actor_combat.obj, and is within the house rule allowing sparse assembly in a math/helper primitive.

Inline ownership

Calling the shared header point_from_line3d directly leaves a candidate-only COMDAT even though January owns none in this object. The production source therefore uses a transparent, coordinate-for-coordinate scalar expansion of that header body. Its semantic local bindings model the inline function's parameters, preserve the exact 0x1240 frame and 0xFC0 padded envelope, and emit no _point_from_line3d symbol or COMDAT. Plain direct statements also remove the COMDAT, but make VC7 collapse the function to 0xFB0; that less faithful schedule was rejected. No forced-inline/noinline attribute, pragma, volatile access, barrier, assembly in the effect routine, or fake dependency is retained.

Final measurements and residual

• Strict gate: 11 exact / 1 residual / 0 unwritten.
• _breakable_surface_effect: target/current padded size 4032/4032.
• Frame: target/current 0x1240/0x1240.
• Relocations: 117/117; relocation types and symbolic destinations occur in the same sequence.
• Freshly regenerated campaign-report fuzzy similarity: 99.8042%.
• Target normalized SHA-256: 510486a7d0a1f2fc1dfb8f0c47990c57f83ce1030ef5e4db8dfc42943dc56bfa.
• Candidate normalized SHA-256: 678a8125d160643a9f1226240128c662bcbb85e7f968f8fd3ed0f0b411f77797.
• All eleven sibling functions and the object's owned data remain strict exact.

The aligned 1,156-target/1,155-candidate instruction diff reduces the genuine code differences to two small x87 regions: commutative operand-role selection in one cross_product3d expansion around target offsets 0x37B-0x392, and an equivalent fxch/load/pop stack schedule in the first scalar-expanded grid offset around 0x97D-0x9B1. More than seventy ordinary-C declaration, scope, association, helper, operand-order, and compiler-shape probes either returned to this fixed point or degraded broader code. Relocation-name differences shown by the historical aligned-diff utility are annotation artifacts; the hardened COFF comparison resolves their symbolic destinations.

Final disposition

This is the most complete authentic, house-rule-compliant source currently available, but it is not byte exact and earns no strict completion credit. It is evidence-parked as instruction-scheduling. Reopen only for authoritative January source/local records, an authenticated compatible compiler/QFE, or a natural same-compiler donor that explains the remaining x87 operand and stack choices.