mirror/BFM-decomp
1
0
Fork 0
mirror of https://github.com/Druthulu/BFM-decomp synced 2026-10-06 01:04:13 -04:00
Code Issues Packages Projects Releases Wiki Activity

docs(phase-29): record the behemoth #2 agent run in the checkpoint

Browse Source 
Opus 5 agent (xHigh) cracking func_8017D960 (3,338 ins, ov_SC03_090). Checkpoint records the
prize (5-member h_norm+h_seq family => ~16,690 distinct-code ins), the MEASURED baseline
(3334 vs 3338, 1806 mismatched, LENGTH-DRIFT/-4) with an explicit correction of the misleading
'one fold OR-chain error left' summary, the sandbox constraints, and the assets it was briefed
with (func_8017F510 matched today = same family at half size; func_8017CA80 matched in the same
TU; §76 allocno-class levers; the -fno-schedule-insns{,2} attribution primitive).
This commit is contained in:
Drew T
2026-07-25 15:15:26 -06:00
parent d8c96a5164
commit a5954802bf
1 changed files with 19 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
phase-ends/CURRENT_PHASE.md
+19 -1
View File
@@ -3765,7 +3765,25 @@ conditional) · main-EXE/B9 + GLM/B6 + resident's 14 walls (P30) · behemoths B7
> **🛑 SESSION-19 CLOSING CHECKPOINT (2026-07-25, Opus 5 @ High) — REFRESHED mid-session; supersedes
> both the SESSION-18 block and the earlier SESSION-19 block (which was written before the
> ENGINE_SHB / class-B / dedup_extend-bug work and went stale). Fresh session safe here.**
> **No background job is running.** Behemoth #3 `func_8017F510` is **CRACKED and BANKED**
> **⚠️ AN OPUS-5 AGENT IS RUNNING (launched 2026-07-25, SESSION-19, effort xHigh):** cracking
> **behemoth #2 `func_8017D960`** (3,338 ins, ov_SC03_090). Deliverables
> `.run/giants/s19_func_8017D960_b5.c` + `s19_d960_report.md`; sandboxed to `.run/giants/`
> (no `src/`/`config/`/`docs/`, no commits, no `make` build targets). **If this session died
> mid-run, check those two files — the agent's work is NOT committed.**
> **THE PRIZE: its h_norm+h_seq family has 5 UNMATCHED members** (`func_8017D960` @ SC03_089/
> 090/104 + `func_8017E778` @ SC03_091 + `func_8017CD9C` @ SC03_102), all §40-remappable from
> one crack ⇒ **≈16,690 distinct-code ins**, ~5.5× both of today's behemoths combined.
> **BASELINE (I measured it): 3334 vs 3338 ins, 1,806 mismatched, `LENGTH-DRIFT/-4`.** The old
> summary "3,334/3,338, one fold OR-chain error left" reads as *4 instructions from done* and
> **is wrong** — it is 4 SHORT, and the drift shifts every downstream position. Length must be
> fixed before any register reasoning (the F510 precedent: dedicated `vw`/`vzw` temps restored
> exact length; a NEGATIVE drift means we are MISSING instructions the target has).
> **Briefed with today's assets:** `func_8017F510` (1,511) MATCHED = the single-light variant of
> the SAME family at half the size, plus `func_8017CA80` (952) MATCHED in the SAME TU — matched
> relatives bracketing the target; cookbook **§76** (allocno class = declaration scope +
> variable reuse), §74 (its 5 pins are audited-safe but F510 cracked PIN-FREE — remove them),
> §77, and the `-fno-schedule-insns{,2}` attribution primitive.
Behemoth #3 `func_8017F510` is **CRACKED and BANKED**
> (97 → MATCH, 1,511 ins, pin-free; §76), and its sibling **`func_8017F5B4` banked mechanically**
> off it (§40 remap + §77 preamble carry, ~0 agent tokens). R22 140/140. The `func_80174CB0`
> sweep is complete and committed.