Files
xenia-rs/audit-runs/audit-030-lr-trace/canary-trace-captures.txt
MechaCat02 8e709b0a24 chore: track audit-runs summary artifacts (md/csv/diff/txt/json/etc)
Snapshot of every non-log artifact under audit-runs/ from audits 003
through 058: findings.md per audit, comparison CSVs, probe diffs,
schema docs, register-dump txts, lr-trace JSONL streams, the saved
canary patch diffs, etc. ~284 files / ~52 MB total.

Excluded (per .gitignore): probe stdout/stderr/log streams (the raw
firehose), guest-memory dumps under audit-026/027/029 (4.5 GB of
.bin files; *.bin pattern added to .gitignore this commit).

Also adds the orphan audit-058-sub825070F0-activation directory that
a subagent accidentally created at project-root instead of
under xenia-rs/audit-runs/; relocated to its proper home.

Purpose: cross-machine continuity. With these summaries committed,
a fresh clone gives the next session the full per-audit context
(findings + tables + cascade predictions) without dependence on
local-only working tree.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-10 21:36:41 +02:00

29 lines
1.7 KiB
Plaintext

# AUDIT-030 canary-side TRACE-PC-LR captures
# Build: xenia-canary (Debug Linux x86_64), 30-LOC patch reverted post-session.
# Patch: src/xenia/cpu/{cpu_flags.{cc,h},backend/x64/x64_emitter.cc,ppc/ppc_hir_builder.cc}.
# Mechanism: --log_lr_on_pc=N emits HIR Trap(100) on each PC=N exec; trap calls
# host TrapLogLR(raw_ctx) which XELOGI's PPCContext.lr + r3..r6, r31.
# ---- Run A: --log_lr_on_pc=0x824D2C08 (sub_824D2C08, audio init) ----
# Wall: 22 s, --disable_instruction_infocache=true, --log_level=3.
# 1 emit message, 1 capture (entry hit during audio init):
i> F8000008 AUDIT-030 emit log-lr trap at HIR pc=824D2C08
i> F8000008 TRACE-PC-LR pc=824D2C08 lr=824D3054 r3=BE628EDC r4=701CF4D0 r5=00000010 r6=00000018 r31=00000000
# Reading: caller is sub_824D2D14-area at lr=0x824D3054. r3 is host pointer
# (0xBE62...) into audio_system v80 memory. r4..r6 are init args.
# ---- Run B: --log_lr_on_pc=0x824D23B0 (sub_824D23B0, alleged per-frame entry) ----
# Wall: 2+ minutes (multiple attempts), --disable_instruction_infocache=true.
# 0 emit messages, 0 captures EVER, even after 21,399 KeReleaseSemaphore(0x828A3230)
# calls + VdSwap=1 in the *first* 2-min run (audio worker active, frame rendered).
# Second 2-min run got stuck at boot due to slow JIT warm-up post-rebuild
# (1190-1240 lines log only). 5-min run also stuck.
# 0/N captures => sub_824D23B0 (PC 0x824D23B0..0x824D2877) is NEVER ENTERED in canary
# during the title-menu loop.
# ---- Verification: --break_on_instruction=0x824D2C08 ----
# SIGTRAP (exit 133) at audio init => break-on-instruction mechanism is live.
# ---- Verification: --break_on_instruction=0x824D23B0 ----
# Process ALIVE for 25 s with audio worker firing => no SIGTRAP =>
# sub_824D23B0 entry not reached, consistent with Run B.