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

antenna.obj exact-match log

Plain-English behavior

This object owns the springy antenna widgets attached to objects. It allocates an antenna datum, builds the chain of simulated vertices from the ant! tag, keeps the root attached to an object marker, advances the spring simulation, and renders each segment as an alpha-blended sprite.

The later HCEA reconstruction is useful as a control-flow and type hint only. Every accepted statement, type, offset, call, constant, and source shape is adjudicated against the January 14, 2002 XDK 3911 object.

Compiler provenance: Microsoft 32-bit C/C++ Optimizing Compiler 13.00.9254.1 from XDK 3911, with the repository's unchanged /O2 /Oy- /DDEBUG /Dxbox flags.

Baseline

• 11 functions total.
• 5 lifecycle/deletion functions were already strict-exact.
• antenna_new, three private helpers, antenna_render, and antennas_update were initially absent.
• Target-owned data consists of two strings; antenna_data is external.

Accepted reconstruction findings

• antenna_datum is 0x2BC; its 21 runtime vertices begin at +0x1C and have a 0x20 stride.
• antenna_definition is 0xD0; its vertices tag block is at +0xC4.
• antenna_vertex_definition is 0x80; its cached vector is at +0x74.
• antenna_render and antennas_update matched exactly from the natural HCEA topology after correcting the January-only final sprite flag to zero.
• antenna_new matches with one inner-scoped real_point3d position, a true short index, the January header-store order, k/j/i velocity clears, position accumulation before the index increment, and one scoped sentinel pointer. This is exact at 0x200, 11 relocations.
• code_001209a0 (attachment refresh) requires standard abs((long)delta) expressions, a short translation-loop index, and a named real_point3d *vertex_position. The pointer shape changes indexed stores from +0x1C/+0x20/+0x24 to the target's +0/+4/+8; this function is now strict-exact at 0x130, 8 relocations.
• code_00120c40 uses a natural pretested for loop with a short index, no count local, and no long-lived vertex-array alias. previous_position and carried_tip live across iterations; the other vectors are iteration-local. The sprite vector has its own nested aggregate-initializer scope, with the angle scoped inside it. This delays the EBX save to the January loop preheader and is exact at 0x310, 15 relocations.
• The simulation also uses a local {0,0,1} up vector, inline cross_product3d, and normalize/fallback to global_left3d.
• code_00120ad0 needs a named long-lived vertices tag-block pointer, an independently indexed next_vertex, and the color copy after all three direction assignments. This is exact at 0x170, 10 relocations.
• Recovered supporting layouts are bitmap_group_sprite 0x20, bitmap_group_sequence 0x40, and build_sprite_data 0xA4; sprite_spacing is signed.

Rejected or superseded shapes

• Direct indexed vertex-position writes in code_001209a0: correct semantics but wrong addressing form; superseded by the named point pointer.
• ABS macro in the jump test: emitted three branch-based absolute-value sequences and oversized the function. The CRT abs intrinsic emits the target cdq; xor; sub idiom.
• Long simulation loop index: cached the definition pointer in EBX, emitted a dword root test, and changed the private helper convention.
• A three-term simulation guard plus do/while compiled to 0x300 and cached the definition pointer in EBX from entry. Adding a named count did not change that allocation.
• A long-lived runtime-vertex pointer restored 0x310 but grew the frame to 0x9C and broke the private helper ABI.
• Manual cross-product expressions: missed two zero-constant relocations and undersized the simulation. The normal inline helper is structurally right.
• Declaration-order-only experiments on antenna_new: moving the three position scalars changed frame slots but did not by itself recover the target pointer/induction topology.

Representative measured topology matrix

Function/shape	Padded size	Relocations	Result
code_00120c40, three-term guard plus do/while	0x300	15	Rejected: definition cached in EBX from entry
code_00120c40, long-lived runtime-vertex pointer	0x310	15	Rejected: frame grew to 0x9C; private ABI drifted
code_00120c40, natural for loop and nested sprite scope	0x310	15	Accepted: strict hash eb6cd6ae...ddd82a
antenna_new, recovered index/pointer lifetime	0x200	11	Accepted: strict hash 27e92e23...0738d
code_001209a0, named vertex-position pointer	0x130	8	Accepted: strict hash 74070256...1d26c
code_00120ad0, tag-block pointer and independent next vertex	0x170	10	Accepted: strict hash 35ec5666...dfd90

Final strict verification

• 11/11 functions strict-exact, totaling 2,432 padded code bytes.
• The 34-byte allocation-error string and 8-byte antenna name string are exact; antenna_data remains external as in January.
• The reconstruction uses XDK 3911 and the symbol-aware hardened comparator. It uses no assembly, volatile, __forceinline, undefined behavior, byte patch, or compiler-flag change.

The object is safe to mark Matching after the consolidated build and whole-object regression gates pass.

2026-08-13 independent salvage revalidation

The completed object was treated as untrusted old work and transplanted from commit 031cf2d67743bee9131347ca51c2be6c3ecdd17b. It was first audited on campaign tip 1bf72e846d3f0b877d19ebfe72243df74bc10dd0, then rebased and fully revalidated on authoritative tip 915c7879c71a70a3381f5179398a2ee848639784. Every target function and every target-owned data section was independently rebuilt and checked with the current hardened comparator after the final rebase.

• All 11 functions are strict-exact: 2,432/2,432 padded code bytes and all relocation addresses, types, destinations, and addends match January.
• Both target-owned strings are exact, including COMDAT ownership/storage: 34 bytes for the allocation error and 8 bytes for the data-array name.
• antenna_data remains an external undefined symbol, matching the target's ownership model; no candidate-owned .data or .bss was introduced.
• bitmap_group.sprite_spacing is proven signed by _antenna_new's January sign-extending load. The first transplant briefly exposed bitmap sequence, sprite, and bitmap-access declarations in the shared bitmap header. Those definitions were localized to antenna.c, because antenna is their only new consumer and keeping the shared header smaller reduces unnecessary compiler and integration blast radius.
• A ten-unit forced-rebuild gate covered every direct consumer of bitmap_group.h, antenna.h, and the new render_sprite.h: antenna, widgets, bitmaps_quantitize, targa_file, tiff_file, predicted_resources, overhead_map, rasterizer_text, rasterizer_xbox_hardware_bitmaps, and rasterizer_xbox_motion_sensor. After regenerating the current objdiff report, every previously accepted function remained exact and there were zero changed_nonexact functions. The only gate failures were the expected Matching-label transition and newly exact antenna evidence.
• The active render_sprite.obj lane changes only render_sprite.c and its log. Its independently recovered build_sprite_data size is also 0xA4, so this new public header neither collides textually nor contradicts the known January layout; the render-sprite lane will consume it only after both atomic objects land.

The house/Berth audit found typed antenna_get, antenna_definition_get, bitmap_group_get, object_get, and TAG_BLOCK_GET_ELEMENT access; one parameter per line; explicit returns in all void functions; and a single return in every function. There is no assembly, volatile qualifier, force-inlining, alignment directive, magic offset, compiler flag change, or comparison exception. The source also labels the original integer-truncation threshold bug and explains the non-matching fabs repair without altering the shipped behavior.

Final authoritative-tip gates

• A clean detached 915c7879 worktree and the antenna lane both force-built all ten direct consumers with XDK 3911. The hardened comparator found every runtime section identical in the other nine objects: code, .rdata, .data, .bss, relocation identity/addends, and ownership all remained unchanged. Debug records were intentionally excluded because their embedded source paths differ between worktrees and are not linked runtime content.
• The pre-change fail-closed regression manifest reported zero changed_nonexact entries for all nine shared-header consumers. Its antenna and configuration findings were the expected reviewed admission transition: six newly implemented functions, newly owned string COMDATs, and the NonMatching to Matching label. A fresh post-admission ten-unit snapshot and rebuild then passed with ok: true, no failures, and no warnings; all 11 antenna functions remained in still_exact.
• ninja halobetacache_build, ninja semantic_progress, and ninja progress completed successfully. The resulting Halo cache progress was 267/468 complete objects, 3,348/7,574 accepted functions, and 406,006/1,770,166 accepted code bytes. Semantic evaluation reported 3,353 accepted-exact functions and zero unit errors.
• The complete tooling suite passed 177/177 unittest tests. The admission audit found no antenna contradiction or revocation, and git diff --check was clean.