From e1e77c7482dcb0e0fab13d2e529c97f3ff20df22 Mon Sep 17 00:00:00 2001 From: Drew T <50529377+Druthulu@users.noreply.github.com> Date: Wed, 22 Jul 2026 00:12:04 -0600 Subject: [PATCH] =?UTF-8?q?docs(phase-29):=20SESSION-7=20CHECKPOINT=20?= =?UTF-8?q?=E2=80=94=20full=20fresh-session=20block=20(=C2=A761c=20is=20th?= =?UTF-8?q?e=20single=20next=20task)?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- phase-ends/CURRENT_PHASE.md | 104 ++++++++++++++++-------------------- 1 file changed, 47 insertions(+), 57 deletions(-) diff --git a/phase-ends/CURRENT_PHASE.md b/phase-ends/CURRENT_PHASE.md index f9bf98db5..8b23efa6d 100644 --- a/phase-ends/CURRENT_PHASE.md +++ b/phase-ends/CURRENT_PHASE.md @@ -902,65 +902,55 @@ conditional) Β· main-EXE/B9 + GLM/B6 + resident's 14 walls (P30) Β· behemoths B7 PLUMBING (Β§57 self-decl et al), not the carve. **Tree reverted; clean-fleet 140/140; nothing banked.** All 12 cracks preserved at `.run/giants/t5wave_*`. Commit `commit:0803`. -> **πŸ›‘ SESSION-6 CHECKPOINT (2026-07-21) β€” safe to open a FRESH session here.** Tree clean (only the R23 -> `db.*.gbf` churn + 4 preserved wave-4 `-O0` drafts). **R22 clean-fleet 140/140 byte-identical**; -> `make tools-health` OK (dedup **1847/0**, C1 234343/234343); 0 NON_MATCHING (G4). HEAD `commit:0772` -> (Drew pushes β€” commits `commit:0771` Task-13A, `commit:0772` Task-13B). +> **πŸ›‘ SESSION-6 CHECKPOINT (2026-07-21, superseded by SESSION-7 below) β€” Task 13A/13B.** +> The classifier + the targeting finding: of the **972** grinder-admissible records only **75 (7.7%)** +> are permuter-shaped, so ~92% of its CPU was unreachable work. Directed run: 24 wins β†’ **21 banked** +> (27%); `func_80141B90` propagated **Γ—138**. Plateau autopsy: **ZERO missing-transforms** β€” the +> plateaus are wrong drafts, so the win was NARROWING admission (permuter bucket 389β†’84), not widening +> the mutation set. Commits `commit:0771`, `commit:0772`, `commit:0796`, `commit:0794`. + +> **πŸ›‘ SESSION-7 CHECKPOINT (2026-07-22) β€” THE CURRENT ONE; safe to open a FRESH session here.** +> Tree clean (only R23 `db.*.gbf` churn + preserved drafts). **R22 clean-fleet 140/140 byte-identical**; +> `make tools-health` OK (dedup 1848/0, C1 234481/234481); 0 NON_MATCHING (G4). HEAD `commit:0804`. +> **Fleet 78.0% instr Β· 66.5% distinct-code Β· 87.9% fn-count** (flat this session β€” see "banked: 0"). +> **Drew pushes** all commits (R6/R20). > -> **THIS SESSION delivered Task 13A + the first directed 13B run.** The headline is not the classifier, it is -> what it MEASURED: of the **972** records the grinder's own filter admits, only **75 (7.7%)** are -> permuter-shaped β€” 547 structural, 348 drafts that are not the function at all. **~92% of the daemon's CPU -> has been going where a search-closer provably cannot win**, which is the byte-grounded explanation of the -> Phase-22 audit's "7 banks all-time, all Phase 21, 0 since." It was never a missing transform; it was -> **targeting**. Fixed free (1,303 β†’ 78 candidates + directed profile from the measured class, since 91% of -> records carry no label and the "directed" search was silently running on gcc defaults). The very first -> directed run then **won on its first candidate** and banked 2. +> **THIS SESSION (after 13A/13B): Task-14 stage 1, Task 5, a 12-agent wave, and the jtbl campaign.** +> * **Task 14 stage 1 β€” the ARITY pre-pass** (`commit:0797`): diagnosed from REAL cc1 stderr (the gate's +> own `memcpy` label is the Β§58 red-herring, 10/11 identical). Banks **2 of 7** vs a **1/12** +> baseline. **NEW HARD CONSTRAINT (Β§61):** any ladder stage mutating SHARED state must undo by +> SNAPSHOT-RESTORE, never an inverse transform, and be verified FLEET-WIDE β€” it broke 138/140 once. +> * **Task 5 β€” Ghidra-C prefetch** (`commit:0798`): `ov_SC06_018` imported, **101 stubs decompiled** +> (cache 882β†’983). Scoping corrected my own claim: the substantial mass is **1,398 h_seq families**; +> **340 (7.72pp) are already-attempted/walled**, **1,058 (5.40pp) fresh**, and the 877 needing a +> prefetch have **median 1 member** (overlay-unique) β†’ prefetch is PER-OVERLAY, one import β‰ˆ 1.59pp. +> * **The wave** (`commit:0799`): 12 agents, ~2M tokens β†’ **11 MATCH / 1 near**, incl. all 3 giants +> (710/673/478). **Gate banked ZERO.** All 12 preserved at `.run/giants/t5wave_*` β€” **DO NOT re-draft.** +> * **jtbl campaign** (`commit:0800`, `commit:0801`, `commit:0803`): the carve-must-follow-splice law +> implemented in `harvest_verify` (temporary splice β†’ carve β†’ isolate β†’ un-splice β†’ re-extract β†’ +> re-derive stub map). `jr_inventory`'s ownership assertion was **RIGHT** β€” it caught MY stranded +> carve. Fixed per-function snapshot-restore; found per-function undo is unsound in a batch +> (isolation repartitions shared source). > -> **β–Ά THE SINGLE NEXT TASK: finish 13B β€” run the directed permuter over the remaining ~73 permuter-bucket -> functions**, then autopsy only what genuinely plateaus. Command: -> `.venv/bin/python tools/grinder.py --once --batch 20 --permute-secs 120 -j 14 --max-closeness 30` -> (repeat; `--no-targeting` A/Bs the filter). Expect ~2 min/fn of CPU, no tokens. **Re-run -> `tools/autopsy.py collect` first** if the backlog has moved (21 s) β€” the corpus is regenerable and -> gitignored by policy. THEN, and only then, the LLM batch autopsy over the survivors: cluster by `sig` -> (`docs/autopsy.md`), one representative per cluster, routing each to missing-transform (extend -> `permuter_weights`) / mis-classified / seed-structural / genuine-wall-with-expiry. +> **β›” THE ONE BLOCKER, AND THE SINGLE NEXT TASK (cookbook Β§61c):** the jtbl carve+isolation path is +> **INCREMENTALLY VALID AND CLEAN-INVALID.** `func_80135A4C` banks every time at the gate +> (BYTE-IDENTICAL) and **fails `make clean && extract-all && check-all` TWICE, identically** +> (139/140, `[FAIL] ov_SC06_018`). The gate cannot see it **because the gate IS the incremental +> build** (Β§42b, worst form). **NEXT: diff the incremental vs clean `build/ov_SC06_018/**` object set +> + the generated `.ld`/asm for the carved subseg. Do NOT bank any jtbl core until it reproduces.** > -> **CARRIED, with measured prices (not projections):** (a) the **306 `integration`** rows β€” a 12-draft gate -> probe banked **1/12** (11 PLUMBING), so this **prices Task 14's Β§57/Β§59 reconcile ladder** rather than -> promising free banks; it is the largest reach-weighted bucket (2,303) and the ladder is the lever. -> (b) the **699 `redraft`** rows are **un-attempted work misfiled as near-misses** (a ~15-ins draft scored -> against a large target; "closeness 278" is a length artefact) β€” fresh crack fuel, and the reason -> `docs/backlog.md`'s closeness ranking overstates how nearly-done the frontier is. (c) `func_80167714` -> (104 ins, reach-134) and the 2 grinder banks are **Γ—1, un-propagated by design** (Β§55b: propagation is its -> own targeted `dedup_propagate --addr` batch). +> **BANKED THIS SESSION: 0.** 21 came from 13B earlier; everything after is tooling + findings. +> Reported as zero, not rounded up. > -> **DO NOT RE-DERIVE (this session's byte-grounded findings):** `residual_class` closeness == -> `masked_diff.structured_diff` on 1,673/1,673 rows, 0 classifier errors (R34 cross-check is baked into -> `autopsy collect` and refuses the corpus on any disagreement). `masked_diff`'s shared probe file was -> losing **0.8% of drafts in every parallel wave** (now per-PID). `gate_stage` never passed `--o0` (phantom -> residuals for every -O0 function, written into this very backlog). The three Β§60a gate defects are fixed. +> **MEASURED, do not re-derive:** of 11 preserved cracks exactly ONE reaches byte-identical via the +> carve path; the other 4 table-bearing ones fail **one-at-a-time too**, on Β§57 plumbing, not the carve. +> The remaining wave drafts' blockers are Β§57 self-decl + local-type redefinition β†’ **Task 14 stages +> 2–3** (`normalize_self_decls` and the type-lift are wired into `family_sweep` but NOT `gate_stage`). > -> **SESSION-5 CHECKPOINT (2026-07-20) β€” superseded by the Task-13A/13B entries above.** Tree clean (only db.*.gbf R23 -> churn), **140/140 byte-identical** (last R22 clean-fleet at the crack-wave-4 verify + func_8017AE2C Γ—137; -> Task-8/12 tooling changes touch only matching-loop tools, no build/src output). tools-health OK (dedup 1846/0 -> class). Main HEAD after this = the checkpoint commit. **Fleet 78.0% instr Β· 66.5% distinct-code Β· 87.9% -> fn-count** (remaining substantial frontier 572 families / 1.58M templatable ins). -> **THIS SESSION delivered (commits `commit:0736`..this):** (1) **crack-waves 1–4** β€” the non-MCP cached substantial -> crack-fuel is now DRAINED; crack-wave 4 = 6 -O2 cores Γ—1 (`commit:0766`), its Γ—138 propagation + 4 -O0 cores -> DEFERRED (integration/-O0 walls). (2) **2 tooling fixes** β€” `family_sweep --source` now honored + `jtbl_family_bank` -> span-fit auto-isolate β†’ **recovered `func_8017AE2C` Γ—137** (`commit:0763`, R22 140/140). (3) **`tools/burndown.py`** -> (the ROI-close instrument, `commit:0764`). (4) **`docs/hindsight-study.md`** (`commit:0765`) β€” the optimal-decomp-ordering -> study for a future Fable roadmap pass; **Β§7 = the permuter-failure-autopsy design**. (5) **Task-12** (`commit:0768`) β€” -> structured failure telemetry NOW FLOWS (`masked_diff.structured_diff` + `match_one --json` + `backlog.residual`/ -> `passes_tried` + `gate_stage` logs the residual per near/fail, verified end-to-end) + grinder durably saves the -> winning C to `.run/permuter-winners/` before banking. -> **THE SINGLE NEXT TASK (fresh session): Task 13 (A+B)** β€” build the deterministic residualβ†’class classifier + -> the LLM-batch autopsy that routes each backlog plateau to missing-transform (extend `permuter_weights` passβ†’class -> map) / mis-classified (fix classifier) / seed-structural (one LLM seed) / genuine-wall (file w/ expiry). Then -> **Task 14 (C+D)** β€” enrich `gate_stage`'s ladder with Β§57/Β§59 so more wins auto-integrate + grinder auto-R22 -> every N wins. Then run the closed loop (improved permuter over the whole backlog β†’ collect telemetry β†’ extend -> ILS β†’ re-run β†’ track the LLM-free fraction via `burndown.py`). **DESIGN SUBSTRATE: `docs/hindsight-study.md` Β§7 + -> Tasks 12–14.** Carried residuals: crack-wave-4's 6 -O2 Γ—138 propagation + 4 -O0 (needs the Task-14 ladder / -O0 -> rollout); the permuter backlog (Task 10, 550 close-1..4 β€” the autopsy loop's input). **Effort: Max for the -> classifier/autopsy design (Task 13 is Tier-1 synthesis); prompt Ultracode only for a breadth backlog-sweep.** -> **Drew pushes all commits (R6/R20).** +> **PHASE STATE:** remaining frontier **22.0pp** (substantial 80–1000 ins = 12.84pp Β· mid 16–79 = +> 8.60pp Β· behemoths 0.28pp Β· small 0.25pp). Remaining tasks: **Β§61c diagnosis (gates all jtbl)** Β· +> Task 14 stages 2–3 Β· Task 4 remainder Β· Task 7 ROI-close Β· ~11 un-propagated small banks Β· the +> βˆ’O0 rollout (splat wall) Β· 2 permanent giant walls. +> **Effort:** xHigh is fine for the Β§61c diagnosis (settled shape); prompt **Max** if it turns into +> non-obvious debugging. **3 stuck self-matching `pgrep` waiters** may still be sleeping (harmless; +> kill by PID if seen).