docs(planning): burst-compute queue — T5500 down, T3.4 fired

.github/workflows/ci.yml

@@ -1,4 +1,6 @@
 # Mirror-only CI config. GitForge (.gitforge.yml) is the pipeline of
+# Keep cargo-deny wildcard checks meaningful for path dependencies by pinning
+# each internal workspace dependency to the workspace package version.
 # record; a red run here is NOT a code-failure signal.
 name: ci-mirror
 

crates/opencodifier-config/Cargo.toml

@@ -9,8 +9,8 @@ authors.workspace = true
 description = "Configuration discovery, precedence, and strict validation for OpenCodifier"
 
 [dependencies]
-opencodifier-core = { path = "../opencodifier-core" }
-opencodifier-engine = { path = "../opencodifier-engine" }
+opencodifier-core = { path = "../opencodifier-core", version = "0.6.1" }
+opencodifier-engine = { path = "../opencodifier-engine", version = "0.6.1" }
 serde = { workspace = true }
 serde_json = { workspace = true }
 thiserror = { workspace = true }

docs/planning/burst-compute-queue-2026-10-26.md

@@ -0,0 +1,83 @@
+# Burst-compute task queue — T5500 down (2026-10-26)
+
+**Trigger:** T5500 powered off at operator direction 2026-10-26 ~1:20 PM
+Central (idle 27B server killed with the box; no training jobs were
+running). Turn-on is operator action, on request.
+
+**Platform policy (binding, Oct 10):** local hosts first; training
+free-tier-first (Kaggle T4); Modal burst ≤ $50 authorized. Cloud
+training stays user-gated per TRAINING.md §8 non-goals.
+
+## What burst compute is NOT needed for (resolved by evidence)
+
+- **Strands consensus pass** — DONE on fedora's 5060 Ti (cu130 venv,
+  0.23 s/row, T1.3 full pass 183,262 rows; T1.4 analysis + T1.5 gate
+  FIRED; merged-v5 built). The 42 GPU-h figure predates the 5060 Ti
+  unlock. Not a burst item.
+- **2B training arms** — T3.3/T3.3' complete (r16/r8 falsified v5;
+  t33p passed both gates). fedora ran these; no burst.
+- **24j 231-split leg** — ONNX verifier, ~49 ms/pass ⇒ minutes on CPU.
+  Engineering item, not burst.
+
+## Burst-relevant queue (ranked; timeframes are wall-clock estimates)
+
+### B1. T3.4 — 4B own-tuned capacity arm [FIRED, top of queue]
+- **Trigger check: FIRED.** Pre-registered trigger = "2B arm plateaus
+  below 0.75 board after v5 + per-kind T." Champion 2B (E1-E) sits at
+  0.6840 board; v5/v6 corpus arms (t33p 0.6580) confirm the plateau.
+  The named contingency applies.
+- **Spec:** Qwen3.8-4B-Distill, r48 LoRA, corpus = e1-iterb scaffold
+  (the only corpus that ever produced a 0.8083-suite arm — v5/v6
+  volume falsified) + T3.2 per-kind calibration; fork-anchor floor
+  0.7662; same release gate (suite ≥ 0.68 AND JevBench ≥ 0.6494).
+- **Host:** Kaggle T4×2 (free, 16 GB each) — 4B r48 LoRA at b4-bucket
+  ckpt ON per the T2.3 sweet spot. Sessions ~9-10 h each (Kaggle
+  12 h cap): epoch 1 ≈ 19k steps × 4B ≈ 30-40 GPU-h total ⇒ **2-3
+  Kaggle sessions across 3-4 days** (weekly 30 h quota). Resume
+  support exists (trainer `--resume`, c44ad6a).
+- **Timeline:** prep (env + corpus upload + smoke) 1 session (day 1);
+  train 2-3 sessions (days 2-5); eval legs + gate 0.5 day (day 6);
+  conditional JevBench leg +0.5 day. **Total ≈ 1 week.**
+
+### B2. JEV-27B contamination-audit probe [needs paid burst or T5500]
+- **Ask:** run `autotrust/JEV-27B` over JevBench-231 through our
+  engine bridge to establish the board-topper's true number on our
+  lane (contamination audit; §14 dominance question).
+- **Host:** 27B inference — NOT Kaggle-T4-able (88 GB class weights
+  at Q3 ≈ 15-16 GB actually; possible on a single T4 at Q3_K but
+  slow ~2-4 s/item ⇒ 231 items ≈ 15 min inference + setup).
+  Realistically: **Modal H100 burst, ~$2-5 of the $50 budget**, 1-2 h
+  wall-clock including image build; or 30 min on T5500 when next
+  powered (the 27B already served there: IQ3_XXS bridge, 2×V100).
+- **Timeline:** T5500-on path = **30 min**; Modal path = **half a day**
+  (image + harness port + run).
+
+### B3. Bridge MTP A/B [T5500-dependent — parked]
+- Draft-mtp is the measured 2× lever on the 27B bridge (fedora memory).
+  The A/B stages (a)/(b) from TRAINING §9 need the T5500 V100 pair.
+- **Timeline when T5500 returns:** half a day.
+
+### B4. Clock-lock sweep + power soak [T5500-dependent — parked]
+- {unlocked,1300,1200,1100} × {120,110,100 W}; PSU-swap-dependent,
+  operator-gated. Half a day when the box returns.
+
+### B5. E1-F/G follow-ons (v3 family remap run) [parked — evidence says no]
+- D42 mechanism shipped; per-family profile thresholds need
+  re-measurement, but the v6 verdict ("confidence-gated volume ≠
+  transfer") downweights any further v3-volume arm. Revisit only if
+  the board submission demands a per-family row.
+
+## Standing next levers NOT in burst scope (engineering, local)
+
+- consensus-flip patch arm — evidence-ranked #1 in TRAINING tail, but
+  it's a 2B fedora arm (9.7 h class), not burst.
+- #150 eval-lane v2 (CPU fork lane determinism tolerance) — local.
+- 24j 231-split leg — local CPU, minutes.
+
+## Turn-on checklist for T5500 (operator)
+
+1. Power on; boot systemd unit applies the 120 W caps (mandatory —
+   PSU OCP regime; `relaunch-campaign.sh` sanitizes torn resume files
+   if any).
+2. Parked items resume: B2 (30 min), B3, B4, and task #18 (Vulkan
+   probe if the V620 has arrived).