Witness/meta probe family + TEKAMOLO resolver + Lo design (surreal_container)

.claude/board/TECH_DEBT.md

@@ -15,6 +15,34 @@
 
 ## Open Debt
 
+### TD-NDARRAY-SIMD-POPCNT-NATIVE — `extract_rules` SIGILLs under `-C target-cpu=native` on larger RowMasks (2026-06-14)
+
+**Surfaced by** the `invariance_witness_probe` (lance-graph-arm-discovery, `--features ndarray-simd`).
+Reproducible: `RUSTFLAGS="-C target-cpu=native"` → the probe prints its header, then dies with
+**signal 4 (SIGILL, exit 132), no Rust panic**, inside `extract_rules` (the per-candidate count loop,
+which under `ndarray-simd` routes RowMasks AND+popcount through `ndarray::simd::U64x8`).
+
+**Characterized:**
+- Crashes at n = 24 000 rows (375 popcount-words) AND n = 24 576 (384 words, no 8-word tail) → **not**
+  a tail-remainder bug; it scales with mask size.
+- The `meta_awareness_probe` at n = 4 096 rows (64 words) runs FINE under the same `native` flags →
+  the faulting kernel is only reached for larger masks.
+- The **default codegen path (no `target-cpu`) and CI's `x86-64-v3` (AVX2, no AVX-512) are unaffected**;
+  the probe result is bit-identical there (INV1 ρ=+0.894 etc.). Both probes are documented to run
+  WITHOUT `native`.
+
+**Most likely cause (unconfirmed):** compile-time `native` on this sandbox enables an AVX-512 feature
+(suspect VPOPCNTQ / AVX-512-VPOPCNTDQ from the U64x8 popcount, possibly AMX) that the **virtualized
+runtime CPU does not actually implement** → illegal instruction. A build-config × sandbox-CPU mismatch,
+not necessarily a logic bug. NOTE: `lance-graph-arm-discovery/Cargo.toml` *recommends* `target-cpu=native`
+/ `x86-64-v4` in a comment — that is the trap a consumer would copy.
+
+**Owed:** simd-savant / sentinel-qa triage to (a) pin the exact faulting instruction (objdump the
+U64x8 popcount kernel + the AMX re-exports), (b) decide whether ndarray's `U64x8` should runtime-detect
+VPOPCNTDQ/AMX rather than trust compile-time `target_feature`, and (c) correct the Cargo.toml comment to
+warn that `native` can SIGILL on virtualized hosts lacking the detected AVX-512 subset. Until then:
+run the `ndarray-simd` probes on the default path. **Status: Open.**
+
 ### TD-CI-COVERAGE-MOLD-1 — `test-with-coverage` job lacks the mold linker the `test` job has (2026-06-12)
 
 **2026-06-12 local-repro addendum (same PR, later commit) — diagnosis CONFIRMED

.claude/knowledge/bf16-hhtl-terrain.md

@@ -3,8 +3,10 @@
 ## READ BY: truth-architect, savant-research, family-codec-smith,
 ##          palette-engineer, integration-lead, container-architect
 
-## STATUS: 5 corrections absorbed, 0 probes run. All claims are CONJECTURES
-##         until the probe queue drains.
+## STATUS: 6 corrections absorbed, 0 probes run. All claims are CONJECTURES
+##         until the probe queue drains. Correction 6 (2026-06-14) reclassifies
+##         the whole float-reconstruction framing as a category error for
+##         byte-encoded models — see the Correction Chain.
 
 ---
 
@@ -42,6 +44,28 @@ Iteration 5: Slot D = CLAM tree path (bucketing > resolution)
   → Correct: Slot D is PRIMARY (bucket address), Slot V is REFINEMENT (exact value)
   → Correct: 3-level 16-way CLAM = 12 bits → 16 → 256 → 4096 alignment
   → UNVERIFIED: requires Probe M1 (CLAM tree fit on 256 Jina centroids)
+
+Iteration 6 (operator, 2026-06-14): float codec on byte-encoded model = CATEGORY ERROR
+  → Measured: bgz-hhtl-d certified on real Qwen3-TTS-1.7B FAILS reconstruction on
+    all 16 roles (palette/hhtld cos ~0.05–0.32, Base17 cos ~0.01–0.37 at the first
+    fold, matvec rel-err ~1.0, fisher-z ρ 0.03–0.68) — cert at
+    .claude/knowledge/certification/hhtld_qwen3tts17b.json. The scalar Slot V (BF16
+    residual) adds ≈0 over palette-only (palette_cos ≈ hhtld_cos).
+  → Root insight: this is NOT just a weak residual. It is a category error. Qwen's
+    information lives in a BYTE-QUANTIZED REGISTER (GGUF / byte-BPE / bgz7), not a
+    continuous float field. Scoring float-RECONSTRUCTION fidelity (Base17 fold +
+    cosine/L2) on a discrete/byte representation measures the wrong category. Same
+    failure mode as I-VSA-IDENTITIES, one layer down: decoding a quantized register
+    back to f32 to score cosine IS the register-loss the iron rule forbids.
+  → Correction: for byte-encoded models the success metric is discrete CODE-AGREEMENT
+    / bucket-ROUTING (the rolling floor), NOT reconstruction. The codec should never
+    decode to f32 to score cosine. Reconstruction is the wrong question.
+  → Evidence the right-category path works (ndarray PR #218 probe family):
+    rolling_floor_probe (coarse code routed via a self-calibrating μ+3σ floor,
+    ρ=1.0 tracking, reject stable ~0.1% under SD drift), helix_bitdepth_probe
+    (direction as a discrete golden index, 24-bit lossless vs ≤f16), and
+    morton_perturbation_probe (discrete frame drives a field on-demand). None
+    decode-to-f32-to-score-cosine.
 ```
 
 ## The Five Hard Constraints

crates/lance-graph-arm-discovery/Cargo.toml

@@ -55,3 +55,23 @@ optional = true
 
 [lib]
 doctest = true
+
+# Meta-awareness probe: real Aerial+ extraction → entropy ladder + Granger meta,
+# witness-as-pointer (meta derived, never stored). Needs the ndarray meta surface,
+# so it rides the `ndarray-simd` feature (the only path that pulls ndarray here).
+[[example]]
+name = "meta_awareness_probe"
+required-features = ["ndarray-simd"]
+
+# Invariance-witness probe: basins as ICP environments — a confounded pair that entropy
+# trusts (low H) but cross-basin invariance refutes (the MIT causal-discovery third axis).
+[[example]]
+name = "invariance_witness_probe"
+required-features = ["ndarray-simd"]
+
+# Coreference rung-separation probe: relative-pronoun resolution over the REAL 144-cell
+# verb_table (VerbFamily × Tense → SlotPrior) as a syntax/Semantik/pragmatic split.
+# Needs the contract (verb_table, via `landing`) AND ndarray (reliability, via `ndarray-simd`).
+[[example]]
+name = "coreference_rung_probe"
+required-features = ["ndarray-simd", "landing"]