Files
Cesarus85--mjolnir-vr/docs/object_matching_logs/errors_obj.md
T

1.9 KiB

source/cseries/errors.obj strategy ledger

This retroactive ledger records preserved evidence for work performed before the mandatory per-object experiment format.

Scope and baseline

  • Compiler: XDK 3911 CL.Exe 13.00.9254.1, /O2 /Oy- /DDEBUG /Dxbox.
  • January target: build/split/source/cseries/errors.obj.
  • Current inventory: 9/10 functions exact; _error is the only residual.
  • Ordinary report: .bss, .data, and .rdata total 501/501 bytes at 100%. Revalidate all three with the hardened section/ownership gate before object admission.

Parked residual

Function Size T/B Relocs T/B Hash T/B Measured residual Class
_error 0x270/0x270 50/50 identities 61bf8aad11f19fd443093f2eba0bae2db27d90af0bc156ec78a530246ae47fd7 / 09377b04504056b8f2278b92bf682d5bd8be32a924e9e87c53cddd954f11ce53 Instruction-identical through +0x111; the tail uses a consistent EBX/ESI/newline-spill allocation split. TU-context optimization

Preserved experiment history

Five grounded families covered declaration order, local scope, register hints, and volatility. Each either emitted the unchanged baseline or regressed relocation addresses. The historical lane did not preserve per-variant hashes or exact source diffs.

volatile is prohibited under current policy even though it appeared in the historical search. Do not repeat declaration/scope/register-hint sweeps without new evidence; never use assembly, byte patches, undefined behavior, or flag changes.

Reopen and disposition

Reopen only with original local-variable provenance, an exact donor explaining the newline spill and EBX/ESI split, or a new legal-C frame/lifetime control that preserves the exact first 0x111 bytes and all 50 relocations.

Disposition: 9/10 code credited; object remains NonMatching, and its data still requires a fresh hardened ownership gate at closeout.