lattice(phase2): H2/H5/H4 verdicts + close CD-H1-01/CD-A2-01 + docs-truth sync

CHANGELOG.md

@@ -7,7 +7,7 @@ All notable changes to ChelatedAI are documented here. Format follows [Keep a Ch
 ### Added
 
 - **Liquified Lattice vision** — [docs/VISION_LIQUIFIED_LATTICE.md](docs/VISION_LIQUIFIED_LATTICE.md) names the north-star (self-annealing pools, shims, evidence DAG, disk scale) with claim boundaries and success metrics.
-- **Phase II execution program** — [docs/ROADMAP_EXECUTION.md](docs/ROADMAP_EXECUTION.md) steps 9–17: Model-Scope close-out, SHIM DoD, annealing controller, evidence DAG, disintegration loop, drift experiment, GNN prototype, quant shim routing, disk pool slice.
+- **Phase II execution program** — [docs/ROADMAP_EXECUTION.md](docs/ROADMAP_EXECUTION.md) steps 9–17. **Delivered apparatus:** rungs 9–12 (Model-Scope close-out, SHIM DoD, annealing controller, evidence DAG); 14 drift experiment (with the H5 living-bank verdict a fail-closed negative); 13 partial (post-bank prune/re-anneal lifecycle only). **Open:** 15 GNN prototype, 16 quant-aware shim routing plane, 17 disk pool slice. The ROADMAP Status column is the source of truth.
 - **Execution queue** — [docs/ROADMAP_EXECUTION.md](docs/ROADMAP_EXECUTION.md) Phase I defines core-first, single-track work (ML → infra → Model-Scope → E2E; SHIM deferred last).
 - **Phase development loop** — `scripts/phase_development_loop.py` registers CORE-SLICE and SHIM-SLICE handlers, persists turn state under `artifacts/phase_loop/`, and emits `NEXT_AGENT_SLICE.json` for the next operator turn.
 - **BHS 10-minute loop tooling** — `scripts/run_10min_priority_bhs_loop.py`, `scripts/loop_core_10m.sh`, `scripts/loop_10m.sh`, `scripts/chelated_loop_timer.py`, and `docs/loop_workers/` worker briefs.
@@ -35,6 +35,11 @@ All notable changes to ChelatedAI are documented here. Format follows [Keep a Ch
 - **SHIM-CD-03 OPEN** — MTP lookahead remains simulation-only (MockMTP).
 - **SHIM-CD-07 OPEN** — BHS program score has not shown quantified lift on §77-83 metrics.
 - **Phase loop** at turn 3171+ (2026-06-05): block flag `CLEAR`, next executable slice `SHIM-SLICE-SCHEDULER-06` (advisory while core queue active).
+- **H5 living-bank VERDICT (closed negative, 2026-07):** on the query-encoder-swap arena the living/annealed post-bank (C5) fails its preregistered gate (C5 must beat **both** C5s frozen-static and C5r one-shot router). SciFact: C5 == C5s (0.131135, bit-identical) and C5r one-shot 0.180862 beats living. NFCorpus: C5 == C5s (0.046389), living edges C5r (0.045817) but does not clear the dual gate. **LIVING BANK WINS = False on both datasets** — the living/annealed lifecycle adds nothing over a frozen static bank; a one-shot router is competitive-or-better. Do not promote the living bank. Sources: [docs/drift-recovery-post-bank-headtohead-results-2026-06.md](docs/drift-recovery-post-bank-headtohead-results-2026-06.md), [docs/drift-recovery-post-bank-headtohead-nfcorpus-results-2026-06.md](docs/drift-recovery-post-bank-headtohead-nfcorpus-results-2026-06.md).
+- **H4 compound-cycles ablation (single-seed):** SciFact C4a seed 42 — `compound_cycles=False` Final NDCG@10 0.236297 vs `True` 0.005258 (~45× collapse toward the frozen floor). Compounding is a rejected design, not a recovery path; the idempotent one-shot fixed point is correct. Source: [docs/drift-recovery-h4-compound-cycles-ablation-2026-07.md](docs/drift-recovery-h4-compound-cycles-ablation-2026-07.md).
+- **H2 swap re-run (post-H1-fix):** SciFact + NFCorpus query-encoder-swap campaigns re-run on the H1-fixed line; regression gate confirms C3a baseline now equals C0/C2/C2O/C4a per seed (contamination was on NFCorpus; SciFact already clean). Closes carried debt CD-H1-01 (stale C3a numbers) and CD-A2-01 (real `all-mpnet-base-v2` swap path exercised). Sources: [docs/drift-recovery-swap-results-2026-06.md](docs/drift-recovery-swap-results-2026-06.md), [docs/drift-recovery-swap-nfcorpus-results-2026-06.md](docs/drift-recovery-swap-nfcorpus-results-2026-06.md).
+- **Lattice rungs 10–14 apparatus (merged, git-verified):** SHIM/promotable-route substrate #284/#285/#286 (+ DoD correction #289, diagnostics observation-only); annealing controller #260 + post-bank temperature schedule #280; Evidence DAG schema #277; disintegration as post-bank prune/re-anneal #279 (+ builder/runtime/conditions #281/#282/#283); concept-drift harness #258–#266 (+ track-0 hygiene #267) with H3 teacher-supervised #287/#288, H4 `compound_cycles` #291, H5 head-to-head driver #290. Rungs **9–12 DONE**, **13 PARTIAL**, **14 apparatus DONE**.
+- **Still OPEN on the Phase II queue:** step **15** GNN prototype (no PyG/DGL code — only a docstring forward-ref in `evidence_dag.py`); step **16** quant-aware shim routing as an integrated steering plane (`adapter_router` + `QuantizationPromotionGate` + retrieval fitness — not closed by #285 alone); step **17** disk pool shard via `block_graph`. Step **13** stays partial until isomer/convergence → Evidence-DAG prune is implemented or its exit criteria are re-scoped to the post-bank lifecycle.
 
 ### Validation (2026-06-06)
 

docs/ROADMAP_EXECUTION.md

@@ -40,7 +40,7 @@ Merge gate: **CLEAR** for this sequence (`docs/next-session.md`). SHIM debts sta
 - No new doc-only shim cycles or 10-agent backlog slices.
 - No claiming SHIM-CD **CLOSED** without full substrate DoD.
 - No Model-Scope + full shim + EGGROLL at full throttle in parallel.
-- No GNN layer or disk-pool integration until Phase II steps 12–14 complete (schema + drift experiment first).
+- No GNN layer or disk-pool integration until Phase II steps 12–14 are honestly closed (schema **#277 DONE**; drift apparatus **DONE** with the H5 living-bank verdict a hard negative; disintegration **#279 PARTIAL** — do not start step 15 while claiming step 13 fully green without a chair re-scope).
 
 ---
 
@@ -51,17 +51,24 @@ Merge gate: **CLEAR** for this sequence (`docs/next-session.md`). SHIM debts sta
 
 Same one-track rule: finish each step before starting the next. Phase II does not reopen parallel doc-only shim cycles.
 
-| Step | Track | Exit criteria |
-|------|--------|----------------|
-| 9 | **Model-Scope shadow (close Phase I #7)** | One bounded steering policy on Qwen3.5-9B fixture; `persist_records` / `load_records` round-trip; bridge `max_total_interventions` exercised in test |
-| 10 | **SHIM substrate DoD** | SHIM-CD-01/02/06 production seams wired without env-only guards; rollback test; restore `Blocking=YES` on open SHIM rows when re-entering |
-| 11 | **Annealing controller** | Single module owns temperature schedule (explore ↔ stabilize); wired to sedimentation + `online_updater` + ES entrypoint; unit test proves high-T increases perturbation, low-T reduces it |
-| 12 | **Evidence DAG schema** | Typed graph contract over `build_attribution_pool.py` output (nodes: query/cluster/actuator; edges: retrieval/intervention links); JSON schema + validator; no GNN required yet |
-| 13 | **Disintegration loop** | `isomer_detector` / `convergence_monitor` triggers prune of low-fitness graph edges or pool entries; re-anneal path records fitness before/after in artifact |
-| 14 | **Concept-drift experiment** | Injected drift fixture (extend `chelatedai-synthetic-collapse` or road-course window); measurable recovery within N anneal cycles documented in `CHANGELOG.md` |
-| 15 | **GNN prototype** | Lightweight GNN over evidence DAG (PyG or DGL); only after steps 12–14 green; must beat flat-pool baseline on drift fixture or fail closed |
-| 16 | **Quant-aware shim routing** | `adapter_router.py` + `QuantizationPromotionGate` integrated as steering plane; promotion requires quant survival + retrieval fitness |
-| 17 | **Disk pool slice** | One precomputed pool shard readable via `computational_storage_poc/block_graph.py` with host parity check; documented in storage track docs |
+**Status snapshot (2026-07, git-verified):** rungs **9–12 DONE**; **13 PARTIAL** (post-bank
+prune/re-anneal lifecycle only — not isomer/convergence → Evidence-DAG prune); **14 apparatus DONE**
+with the **H5 living-bank question closed as a hard negative** (LIVING BANK WINS = False on SciFact +
+NFCorpus); **15–17 OPEN**. Lattice apparatus PRs: #260, #277, #279–#291. Next executable feature work
+is step **15** only after chair re-scopes or closes the step **13** remainder — or **16/17** if the
+quant plane / disk pool is prioritized over GNN.
+
+| Step | Track | Status | Exit criteria |
+|------|--------|--------|----------------|
+| 9 | **Model-Scope shadow (close Phase I #7)** | **DONE** (fixture path; Model-Scope stack + #254 persist/cap tests) | One bounded steering policy on Qwen3.5-9B fixture; `persist_records` / `load_records` round-trip; bridge `max_total_interventions` exercised in test |
+| 10 | **SHIM substrate DoD** | **DONE** (rung 10: #284 A1a / #285 A1b / #286 A1c; DoD honesty #289 — diagnostics observation-only; live routes overlap rung 16) | Promotable steering-route control plane, default-safe, with quant-survival + actionable rollback (`docs/rung10-shim-substrate-dod.md`). This is the rung-10 DoD, **not** a claim that historical SHIM-CD-01/02/06 rows are CLOSED. |
+| 11 | **Annealing controller** | **DONE** (#260 engine controller + #280 post-bank schedule; schedule ownership split across two modules) | Temperature schedule(s) drive explore ↔ stabilize; unit tests prove high-T vs low-T behavior; engine wires the controller into sedimentation temperature; post-bank path uses `annealing_schedule` in the C5 lifecycle |
+| 12 | **Evidence DAG schema** | **DONE** (#277 `evidence_dag.py` + validator + JSON schema; no GNN) | Typed graph contract over `build_attribution_pool.py` output (nodes: query/cluster/actuator; edges: retrieval/intervention links); JSON schema + validator; no GNN required yet |
+| 13 | **Disintegration loop** | **PARTIAL** — post-bank prune/re-anneal DONE (#279 H5a, wired #281–#283); **NOT** isomer/convergence → Evidence-DAG edge prune | Delivered: fitness-gated prune + re-anneal on `SteeringPostBank` with lifecycle artifacts. Original exit named `isomer_detector` / `convergence_monitor` triggers on DAG edges/pool entries — **not implemented**. Re-scope exit to the post-bank mechanism, or implement that trigger. |
+| 14 | **Concept-drift experiment** | **DONE apparatus** (#258–#266 harness, #267 track-0 hygiene; swap arena #268–#276; H3 #287/#288; H4 #291; H5 driver #290). **H5 living-bank VERDICT: FAIL / non-promoted** (SciFact + NFCorpus) | Injected drift + recovery campaigns under `docs/drift-recovery-*.md`. Living annealed post-bank (C5) does **not** beat the frozen static bank + one-shot router gate. Compounding (`compound_cycles=True`) is catastrophic on the single-seed H4 ablation. |
+| 15 | **GNN prototype** | **OPEN** (no merged PR; no PyG/DGL code — only a docstring forward-ref in `evidence_dag.py`) | Lightweight GNN over evidence DAG (PyG or DGL); only after steps 12–14 green; must beat flat-pool baseline on drift fixture or fail closed |
+| 16 | **Quant-aware shim routing** | **OPEN** (pieces exist: `adapter_router`, `QuantizationPromotionGate`, route gate #285 — **not** integrated as one retrieval-fitness steering plane) | `adapter_router.py` + `QuantizationPromotionGate` integrated as steering plane; promotion requires quant survival + retrieval fitness |
+| 17 | **Disk pool slice** | **OPEN** (no pool-shard parity via `block_graph`; the computational-storage POC is a different graph) | One precomputed pool shard readable via `computational_storage_poc/block_graph.py` with host parity check; documented in storage track docs |
 
 ### Phase II dependencies (do not skip)
 

docs/drift-recovery-h4-compound-cycles-ablation-2026-07.md

@@ -0,0 +1,20 @@
+# Drift Recovery — H4 compound-cycles ablation — July 2026
+
+Arena: `query_encoder_swap`. Task SciFact, condition **C4a** (supervised unbounded adapter),
+seed 42, cycles 12, swap `all-MiniLM-L6-v2 → all-mpnet-base-v2`, correction 30 steps @ lr 0.01.
+
+Question (H4 "make cycles compound"): does re-training the adapter from accumulated weights and
+applying each correction to the already-mutated store (`compound_cycles=True`) improve recovery over
+the idempotent one-shot fixed point (`compound_cycles=False`)?
+
+| `compound_cycles` | Final NDCG@10 (seed 42) | Artifact |
+|---|---:|---|
+| False (idempotent fixed point) | **0.236297** | `experiment_runs/drift-recovery/h4-compound/C4a_compound0_seed42.json` |
+| True (compounding) | **0.005258** | `experiment_runs/drift-recovery/h4-compound/C4a_compound1_seed42.json` |
+
+**Result: compounding is catastrophically worse — recovery collapses ~45× (0.236 → 0.005), to near the
+frozen floor.** This quantifies the paper §5.2 "overshoot" observation: applying each bounded
+correction to the already-mutated store escalates the per-cycle correction norm cycle-over-cycle
+(unstable drift, not convergence), so the idempotent re-initialization from a fixed seed against the
+frozen pre-drift snapshot is the correct, stable design — not a failure to compound. This is a
+single-seed ablation (seed 42); the direction is unambiguous but the exact magnitude is one seed.

docs/drift-recovery-post-bank-headtohead-nfcorpus-results-2026-06.md

@@ -0,0 +1,26 @@
+# Drift Recovery — Post-Bank Head-to-Head (H5) — June 2026
+
+Arena: `query_encoder_swap`. Task NFCorpus, cycles 12, seeds [42, 1337, 7].
+
+Conditions: C0 frozen (floor) · C2O oracle (ceiling) · **C5 living bank** · **C5s frozen static bank** · **C5r one-shot router**. All scored on the SAME eval subset.
+
+## Head-to-head (mean over seeds)
+
+| Condition | Baseline NDCG | Final NDCG | Std | Applied runs |
+|---|---:|---:|---:|---:|
+| C0 | 0.597642 | 0.040655 | 0.014475 | 0/3 |
+| C2O | 0.597642 | 0.611933 | 0.018032 | 0/3 |
+| C5 | 0.597642 | 0.046389 | 0.005785 | 3/3 |
+| C5r | 0.597642 | 0.045817 | 0.005607 | 3/3 |
+| C5s | 0.597642 | 0.046389 | 0.005785 | 3/3 |
+
+## Verdict (the H5 gate)
+
+- C5 living mean: 0.04638866344039569
+- C5s static mean: 0.04638866344039569
+- C5r one-shot mean: 0.04581675856141487
+- Living beats static (C5 > C5s): **False**
+- Living beats one-shot (C5 > C5r): **True**
+- **LIVING BANK WINS (beats both): False**
+
+Gate: the living annealed bank must beat BOTH the frozen static bank and the one-shot router with the lifecycle exercised. This verdict is from the real run; it is not asserted until the campaign executes on the GPU.

docs/drift-recovery-post-bank-headtohead-results-2026-06.md

@@ -0,0 +1,26 @@
+# Drift Recovery — Post-Bank Head-to-Head (H5) — June 2026
+
+Arena: `query_encoder_swap`. Task SciFact, cycles 12, seeds [42, 1337, 7].
+
+Conditions: C0 frozen (floor) · C2O oracle (ceiling) · **C5 living bank** · **C5s frozen static bank** · **C5r one-shot router**. All scored on the SAME eval subset.
+
+## Head-to-head (mean over seeds)
+
+| Condition | Baseline NDCG | Final NDCG | Std | Applied runs |
+|---|---:|---:|---:|---:|
+| C0 | 0.818737 | 0.006394 | 0.005594 | 0/3 |
+| C2O | 0.818737 | 0.813164 | 0.003981 | 0/3 |
+| C5 | 0.818737 | 0.131135 | 0.009802 | 3/3 |
+| C5r | 0.818737 | 0.180862 | 0.009034 | 3/3 |
+| C5s | 0.818737 | 0.131135 | 0.009802 | 3/3 |
+
+## Verdict (the H5 gate)
+
+- C5 living mean: 0.13113546862429465
+- C5s static mean: 0.13113546862429465
+- C5r one-shot mean: 0.18086182834701145
+- Living beats static (C5 > C5s): **False**
+- Living beats one-shot (C5 > C5r): **False**
+- **LIVING BANK WINS (beats both): False**
+
+Gate: the living annealed bank must beat BOTH the frozen static bank and the one-shot router with the lifecycle exercised. This verdict is from the real run; it is not asserted until the campaign executes on the GPU.

docs/drift-recovery-swap-nfcorpus-results-2026-06.md

@@ -1,14 +1,7 @@
 # Drift Recovery — Query-Encoder-Swap Arena (PR-A4) — June 2026
 
-> ⚠️ **C3a rows superseded — pending re-run (CD-H1-01).** The C3a baseline and final
-> NDCG below were generated before the H1 harness fix (commit `5b2379b`), which removed a
-> build-time bounded adapter that had contaminated C3a's ingested doc vectors and its
-> pre-correction baseline. After the fix C3a's baseline equals the other conditions
-> exactly and its final NDCG (and the C3a budget sweep) shift slightly; those C3a numbers
-> will be regenerated by the H2 re-run. **C0/C2/C2O/C4a are unaffected.**
-
 Arena: `query_encoder_swap` (encoder upgrade; the C2 re-embed oracle is defeated).
-Manifest: `experiment_runs\drift-recovery\swap-nfcorpus/swap-campaign-manifest-2026-06.json`.
+Manifest: `experiment_runs/drift-recovery/swap-nfcorpus/swap-campaign-manifest-2026-06.json`.
 Base model: `sentence-transformers/all-MiniLM-L6-v2`; swap model: `all-mpnet-base-v2`.
 Task NFCorpus, anchor_fraction 0.4, cycles 12, seeds [42, 1337, 7].
 
@@ -21,22 +14,20 @@ Conditions: C0 frozen (lower bound) · C2 re-embed-with-original (proven no-op)
 | C0 | 0.597642 | 0.040655 | 0.014475 | 0/3 | 0/3 |
 | C2 | 0.597642 | 0.040655 | 0.014475 | 0/3 | 0/3 |
 | C2O | 0.597642 | 0.611933 | 0.018032 | 3/3 | 0/3 |
-| C3a | 0.598189 | 0.053501 | 0.008779 | 0/3 | 3/3 |
-| C4a | 0.597642 | 0.054347 | 0.004909 | 0/3 | 3/3 |
+| C3a | 0.597642 | 0.052435 | 0.012268 | 0/3 | 3/3 |
+| C4a | 0.597642 | 0.052922 | 0.005116 | 0/3 | 3/3 |
 
 ## C3a Training-Budget Sweep (seed 42)
 
 | Steps | LR | Final NDCG | should_correct | applied | mean norm |
 |---:|---:|---:|---:|---:|---:|
-| 30 | 0.01 | 0.051264 | 12/12 | 12/12 | 0.515708 |
-| 200 | 0.05 | 0.039031 | 12/12 | 12/12 | 0.506980 |
-| 1000 | 0.10 | 0.040058 | 12/12 | 12/12 | 0.508960 |
-| 2000 | 0.10 | 0.041228 | 12/12 | 12/12 | 0.509369 |
+| 30 | 0.01 | 0.040350 | 12/12 | 12/12 | 0.515658 |
+| 200 | 0.05 | 0.038064 | 12/12 | 12/12 | 0.507013 |
+| 1000 | 0.10 | 0.034303 | 12/12 | 12/12 | 0.509030 |
+| 2000 | 0.10 | 0.038662 | 12/12 | 12/12 | 0.509522 |
 
-Best budget cell: steps=30, lr=0.01 (seed-42 final 0.051264). Larger budgets degrade (all 12/12 corrections applied at every budget level — the degradation is genuine overshooting, not a broken loop). Three-seed confirmation:
+Best budget cell: steps=30, lr=0.01 (seed-42 final 0.040350). Three-seed confirmation:
 
 | Steps | LR | Mean final NDCG | Std | Recovery@N |
 |---:|---:|---:|---:|---:|
-| 30 | 0.01 | 0.053501 | 0.008779 | 0/3 |
-
-Note: supervised recovery (C3a/C4a) saturates at cycle 1 and holds flat across all 12 cycles. Subsequent cycles re-apply the same bounded correction magnitude without further improvement. Wall-clock timing was not captured (split run); see campaign-completion.log for session timestamps.
+| 30 | 0.01 | 0.052435 | 0.012268 | 0/3 |