docs(pr-x12): address Codex + CodeRabbit review on #242

- tests/knowledge/test_pr_x12_knowledge_docs.py: the four assertions
  pinning the audit-flagged wording (tropical_spmv free function,
  'Actual kernel home (current)' inversion) now assert the corrected
  canon instead — blasgraph canonical (kernel unwritten), shipped
  min-plus = the method bgz17::ScalarCsr::spmv_min_plus. Suite green
  (117 tests). [Codex P2]
- Leaf-count arithmetic corrected everywhere: 129,600 exact 8x8 leaves
  at 4K (padded 130,560; the old 132,710 was ungrounded), ~129 ns/leaf
  budget, RDO paragraphs re-based on ~2,040 CTUs/frame, R-11 matrix row
  re-unitized to <=210 ns/leaf. [CodeRabbit]
- Determinism claims qualified: standardized NN-in-loop decode can be
  bit-exact (fixed-point conformance), so the moat is stated as
  reproducibility-by-construction + governance, not 'NN =
  nondeterministic'; un-standardized AI-upscaler drift claim retained.
  [CodeRabbit]
- H.267 described as prospective standardization effort; VVC metric
  split (up to ~50% subjective / ~31-40% objective BD-rate); GEMM-lens
  'all four already exist' leftover removed; tropical-GEMM speedup made
  conditional on the A6 bench; MD028 blockquote fix in the two
  quarantine headers; residual affirmative stale symbols corrected in
  pr-x12-cam-pq-sigker-dn-tree-substrate-bindings.md. [CodeRabbit]

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01K3RyLEbuNSHxxB3NTTrGki

Author: claude

.claude/PR-X12-docs-audit.md

@@ -340,7 +340,13 @@ the audited docs"). The following recommendations were executed on branch
   **[UNCALIBRATED ESTIMATES]** in delta §6, substrate-canon R-5, GEMM lens
   §2.2; the self-fabricating "Graviton=128" withdrawn from mapping §5.3.
 - Finding #8 — "132,710 CTUs/frame" mislabel: fixed to leaves-at-8×8 in
-  delta §2.3 (budget recomputed per-leaf).
+  delta §2.3 (budget recomputed per-leaf). Review follow-up (same day):
+  the count itself was also wrong — exact 8×8 accounting at 3840×2160 is
+  **129,600** leaves (padded 64×64: 2,040 CTUs → 130,560); the audit's own
+  finding #8 had accepted 132,710 as "correct-as-leaves," which it wasn't.
+  Budget recomputed to ~129 ns/leaf; the "~132K CTUs" figures in the three
+  RDO paragraphs corrected to ~2,040 CTUs; R-11 matrix row re-unitized to
+  ≤210 ns/leaf.
 - Finding #9 — false `signature_kernel_pde` Goursat-PDE bug claim:
   withdrawn at all four sites (delta §11/§12, substrate-canon R-14/R-15).
 - Finding #19 — §9 falsifiability matrix: tagged FORWARD-CONDITIONAL

.claude/blackboard.md

@@ -690,7 +690,9 @@ default is x86-64-v3 (avx2) so ndarray_simd activates on avx512 builds only.
   method `ScalarCsr::spmv_min_plus`, lossy sibling); blasgraph restored as
   bit-exact canon over bgz17; per-arch DCT crossovers tagged [UNCALIBRATED];
   false `signature_kernel_pde` Goursat-bug claim withdrawn (its convergence
-  tests pass); 132,710 = leaves-at-8×8 not CTUs; §9 falsifiability matrix tagged
+  tests pass); the R-11 unit is leaves-at-8×8 not CTUs, and the count itself
+  was corrected to 129,600 exact (padded 130,560; the old 132,710 was
+  ungrounded) → ~129 ns/leaf budget; §9 falsifiability matrix tagged
   FORWARD-CONDITIONAL. Tier-1 docs (`woa-multiarch-orchestration`,
   `bgz-jc-substrate-synergies`) ⛔ QUARANTINED pending rewrite.
 - **NEW: `.claude/knowledge/pr-x12-h266-h267-standards-landscape.md`** —

.claude/knowledge/pr-x12-bgz-jc-substrate-synergies.md

@@ -12,8 +12,7 @@
 > **siblings**, never substitutes for blasgraph. **Do not cite §0 / §2.5 /
 > §2.7 / §4 / §7-§8 as evidence.** Pending rewrite per audit Tier-1
 > recommendation.
-
-
+>
 > Date: 2026-05-22
 > Status: **substrate grounding doc** — connects PR-X12's abstract substrate claims to the **already-implemented** crates in `lance-graph/crates/`. Companion to the five perspective lenses written 2026-05-22.
 >

.claude/knowledge/pr-x12-cam-pq-sigker-dn-tree-substrate-bindings.md

@@ -284,7 +284,7 @@ Updating the inventory from `pr-x12-bgz-jc-substrate-synergies.md` §7 with the
 | 4096-entry basin codebook | `bgz-tensor::Codebook4096` (literal 4096-entry type), trained by **`cam_pq`**. `bgz-hhtl-d` is a *different* basin-codebook strategy (4-basin × 16-HIP × 256-TWIG = 16,384-cell address space over a shared 256-entry palette) — not the canonical 4096 |
 | `CurveOrder<const N>` | `highheelbgz` spiral addressing |
 | `LinearReduce<T> + Basis<T>` | `bgz-tensor` AttentionSemiring + ComposeTable + DistanceTable; **`sigker::SignatureBasis`** (proposed) |
-| Tropical-GEMM (R-7) | `bgz17::scalar_sparse::tropical_spmv` |
+| Tropical-GEMM (R-7) | canonical home `lance-graph::blasgraph` (kernel unwritten); shipped min-plus is the method `bgz17::ScalarCsr::spmv_min_plus` [lossy sibling, prototype only — corrected 2026-07-16 per audit] |
 | Federated codebook (R-13) | `bgz-hhtl-d` shared-palette + **`cam_pq::CamCodebook`** + **`dn_tree`** (online update) + **`merkle_tree`** (integrity) |
 | Formal correctness — codec quantization | `jc` **Pillar 10 (Pflug-Pichler)** — nested-distance Lipschitz on Sigma DN-trees, certifies CAM-PQ tree quantization preserves FreeEnergy within Lε |
 | Formal correctness — path-signature lane | `jc` **Pillar 11 (Hambly-Lyons)** via **`sigker`** — certifies Index-regime classification (sigker only, not bgz) |
@@ -335,15 +335,22 @@ Recommended edits to `pr-x12-canon-resolutions-delta.md`:
 > - **Quantization correctness (Pillar 10, Pflug-Pichler):** nested-distance Lipschitz on Sigma DN-trees — proves CAM-PQ tree quantization preserves FreeEnergy within Lε. This is the proof PR-X12 cites for "wire-format quantization is faithful."
 > - **Path-signature correctness (Pillar 11, Hambly-Lyons):** signature uniqueness on tree-quotient — proves any path is uniquely determined by its truncated signature up to tree-like equivalence. Active under `--features hambly-lyons` (since 2026-05-07, PR #348). This is the proof PR-X12 cites for the `SignatureBasis<DEPTH>` lane (R-15).
 >
-> Both pillars exist; the codec cites them and does not reprove. **Status: Pillar 10 active; Pillar 11 active under feature gate. Production-scale benchmarking + PR #350 (signature_kernel_pde math correction) — see Gap G-4.**
+> Both pillars exist; the codec cites them and does not reprove. **Status: Pillar 10 active; Pillar 11 active under feature gate. Production-scale benchmarking — see Gap G-4.** *(Corrected 2026-07-16, audit #9: the "PR #350 signature_kernel_pde math correction" claim is withdrawn — the PDE form's convergence tests to `I₀(2·√⟨u,v⟩)` pass; there is no known bug.)*
 
-**R-7 path correction** — the kernel home:
+**R-7 kernel home** *(corrected 2026-07-16, audit #1-#4 — the earlier "path
+correction" here had the canon/adapter relationship inverted)*:
 
-> R-7 (corrected): tropical-GEMM lives at `lance-graph::bgz17::scalar_sparse::tropical_spmv` (not the abstract `blasgraph` namespace). The codec's tropical-GEMM RDO call is `bgz17::scalar_sparse::tropical_spmv(edge_weights, dag)`.
+> R-7 (corrected): the canonical, bit-exact home for the tropical-GEMM
+> partition kernel is `lance-graph::blasgraph`; that f32 min-plus kernel is
+> UNWRITTEN today. The only shipped min-plus is the method
+> `bgz17::ScalarCsr::spmv_min_plus` (`fn(&self, x: &[f32]) -> Vec<f32>`) —
+> a lossy-sibling prototype, never a substitute for the blasgraph canon.
+> The free function `tropical_spmv(edge_weights, dag)` cited previously
+> does not exist.
 
 **R-15 (new candidate)** — signature-basis as Basis<T> impl:
 
-> R-15 (candidate): the substrate supports path-structured signals via `sigker::SignatureBasis<DEPTH>: Basis<f32>`, alongside `DctIIBasis<N>: Basis<i16>` (video) and `EwaSplatBasis: Basis<f16>` (3DGS). Implementation: ~1 week wrapper around `sigker::signature_kernel_pde`. **Plan G** gets a fifth lane (path-structured: audio waveform, time-series, gesture/handwriting).
+> R-15 (candidate): the substrate supports path-structured signals via `sigker::SignatureBasis<DEPTH>: Basis<f32>`, alongside `DctIIBasis<N>: Basis<i16>` (video) and `EwaSplatBasis: Basis<f16>` (3DGS). Implementation: ~1 week wrapper around `sigker::signature_truncated` (the form Pillar 11 cites; the PDE form is equally sound — audit #9 — but the truncated path is what R-15 commits). **Plan G** gets a fifth lane (path-structured: audio waveform, time-series, gesture/handwriting).
 
 ---
 

.claude/knowledge/pr-x12-canon-resolutions-delta.md

@@ -159,11 +159,12 @@ Three-way pass per load: (ratio + quality + LoC). Sub-threshold on any one = blo
 |---|---|
 | 4K resolution | 3840 × 2160 = 8.3 M pixels |
 | 60 fps | 16.67 ms/frame |
-| 8×8 leaf | 132,710 leaves/frame (≈2,025 CTUs/frame at 64×64) |
-| **Per-leaf budget** | **125 ns/leaf** |
+| 8×8 leaf | 129,600 leaves/frame (exact: 3840·2160/64; padded 64×64 accounting: 2,040 CTUs → 130,560 leaves) |
+| **Per-leaf budget** | **~129 ns/leaf** (16.67 ms / 129,600) |
 
-> Corrected 2026-07-16 (audit #8): the 132,710 figure counts 8×8 **leaves**, not
-> 64×64 CTUs — substrate-canon R-11 labels it correctly; this table had mislabeled it.
+> Corrected 2026-07-16 (audit #8 + review): the unit is 8×8 **leaves**, not
+> 64×64 CTUs, and the previously-cited 132,710 count was not numerically
+> grounded (exact 8×8 accounting gives 129,600; padded 64×64 gives 130,560).
 
 Encoder per-leaf breakdown:
 
@@ -177,7 +178,7 @@ Encoder per-leaf breakdown:
 | rANS encode (A7) | ~40 ns | ~40 ns |
 | **Total** | **~960 ns** | **~210 ns** |
 
-Scalar misses 60 fps by 7.6×; SIMD-batched misses by 1.7× (same OoM). **Pins B:D-CODEC-8 / A:T-7 from P2 → P1** — A4-impl and A6 must ship SIMD-batched, not scalar-then-vectorize.
+Scalar misses 60 fps by ~7.5×; SIMD-batched misses by ~1.6× (same OoM). **Pins B:D-CODEC-8 / A:T-7 from P2 → P1** — A4-impl and A6 must ship SIMD-batched, not scalar-then-vectorize.
 
 ---
 
@@ -218,7 +219,7 @@ Tropical-semiring (+, min) formulation:
 
 **Complexity:** `O(d² × |nodes|)`. For d=4, |nodes|=85: 1360 ops/CTU vs 21,760 naive. **~16× speedup.**
 
-At 4K 132K CTUs/frame: ~4 ms vs ~64 ms just for partition RDO. At 60 fps, the difference between fitting and missing budget.
+At 4K (~2,040 CTUs/frame; corrected 2026-07-16 — "132K" was the leaf count, not CTUs): ~2.8 ms vs ~44 ms just for partition RDO at ~1 op/ns. At 60 fps, the difference between fitting and missing budget.
 
 **Dep direction:** `ndarray-codec → lance-graph::blasgraph` (tropical-GEMM kernels nominally live in blasgraph). Allowed post-Plan-H because ndarray-codec is a sibling crate, not the bottom.
 
@@ -386,7 +387,7 @@ Highlights of falsifiers — the canary tests:
 | R-1 | A7 has to subclass `LinearReduce` to make rANS work | Trait factoring wrong; A7 wastes 1.5 wks |
 | R-3 | Cumulative generic LoC > 1500 after A4-A8 | M:H-NEW-2 falsified; the abstraction grew domain-specific code |
 | R-9 | `grep -E 'North|East|West|South' src/hpc/codec/*.rs` returns production paths | Topology-free contract broken; consumer semantics leaked into codec |
-| R-11 | SIMD-batched encode > 210 ns/CTU on SPR | Plan G video threshold can't pass; 4K real-time falsified |
+| R-11 | SIMD-batched encode > 210 ns/leaf on SPR | Plan G video threshold can't pass; 4K real-time falsified |
 
 ---
 

.claude/knowledge/pr-x12-codec-cognitive-substrate-mapping.md

@@ -522,7 +522,7 @@ Per-block butterfly wins for single 32×32. Per-frame batched `C = A_batch @ DCT
 
 > [Pinned as **R-7**: the tropical-GEMM kernel's canonical home is `lance-graph::blasgraph` — but the symbol `blasgraph::tropical_gemm` **does not exist** (audit #3, corrected 2026-07-16; blasgraph's 7 HDR semirings are binary-Hamming over 16384-bit BitVec, no numerical min-plus). The kernel is unwritten; the only shipped min-plus is `bgz17::ScalarCsr::spmv_min_plus` (lossy sibling, prototype only). The `ndarray-codec → lance-graph` dep direction was confirmed *allowed* post-merge (both are sibling crates above `ndarray::hpc` and below `woa-rs`). See the corrected R-7 in the delta doc.]
 
-x265 spends ~30% CPU on recursive partition RDO. Reformulate: each partition is a node in an 85-node DAG, edges = split/merge transitions, weights = ΔRDO. Optimal partition = shortest path. blasgraph's tropical-semiring GEMM (`D ← min(D, D + W)`) solves all partitions in **one batched matrix-relax**. `O(4^d)` → `O(d²)` per CTU.
+x265 spends ~30% CPU on recursive partition RDO. Reformulate: each partition is a node in an 85-node DAG, edges = split/merge transitions, weights = ΔRDO. Optimal partition = shortest path. A **planned** tropical-semiring GEMM in blasgraph (`D ← min(D, D + W)`; unwritten today, per the R-7 correction above) would solve all partitions in one batched matrix-relax — the `O(4^d)` → `O(d²)` reduction holds algebraically but the speedup is **conditional on the future A6 partition bench** confirming it on real edge weights.
 
 ### 13.4 CABAC context modeling → tiny transformer (E-9)
 

.claude/knowledge/pr-x12-h266-h267-standards-landscape.md

@@ -92,12 +92,19 @@ now supplies evidence for both directions:
 - *For NNVC:* every single CfE response for H.267 incorporated some form of
   NN tool — the field has voted.
 - *For the inversion:* the decoder-complexity/power pushback (Samsung et
-  al.) is precisely the failure mode the anti-neural rule predicts, and the
-  determinism argument (no model-version drift, legally/scientifically
-  defensible reconstruction — x266 lens §7) is one no NN-in-loop codec can
-  make.
-
-This is the sharpest external differentiator the PR-X12 line has. Keep it.
+  al.) is precisely the failure mode the anti-neural rule predicts. The
+  reproducibility argument needs stating precisely: a *standardized*
+  NN-in-loop codec can be bit-exact (JVET conformance requires fixed-point
+  decode), so "NN = nondeterministic" is wrong as a universal claim. The
+  PR-X12 advantage is **reproducibility-by-construction and governance**:
+  closed-form math with no model artifact to pin, version, or govern —
+  versus NN tools whose reproducibility depends on disciplined model
+  pinning and whose behavior is not analytically auditable (x266 lens §7).
+  Un-standardized AI-upscaler pipelines (DLSS-class, RIFE-class) DO drift
+  across model versions; that comparison stands unqualified.
+
+This — stated in the qualified form above — is the sharpest external
+differentiator the PR-X12 line has. Keep it, and keep it precise.
 
 ## 4. H.267 — the process, the requirement, the dates
 

.claude/knowledge/pr-x12-substrate-canon-resolutions.md

@@ -545,9 +545,10 @@ blasgraph`. For d=4, |nodes|=85: O(16 × 85) = O(1360) ops per CTU.
 Vs. O(4^4 × |nodes|) = O(21,760) ops for the naive recursive RDO.
 ```
 
-**Speedup: ~16×.** For a 4K frame at ~132K CTUs, this is the difference
-between ~4 ms and ~64 ms per frame just for partition RDO. At 60 fps,
-that's the difference between fitting and missing the latency budget.
+**Speedup: ~16×.** For a 4K frame at ~2,040 CTUs (corrected 2026-07-16 —
+the earlier "~132K CTUs" was the 8×8 leaf count), this is the difference
+between ~2.8 ms and ~44 ms per frame just for partition RDO at ~1 op/ns.
+At 60 fps, that's the difference between fitting and missing the budget.
 
 **Why this targets `lance-graph::blasgraph`:** Standard BLAS GEMM uses
 (× , +) semiring. Tropical uses (+ , min) semiring. blasgraph is the
@@ -796,12 +797,13 @@ to the budget.
 60 fps = 16.67 ms/frame
 At 8×8 leaf granularity (HEVC's smallest CU; the unit at which the
 encoder's inner-loop work is paid):
-                              132,710 leaves/frame
-                              (= 2,040 CTUs/frame at 64×64, × ~64
-                               leaves/CTU at maximum split depth;
-                               130,560 from clean 3840·2160/64, with
-                               ~1.6 % bias for chroma alignment)
-Per-leaf budget: 16.67 ms / 132,710 = 125 ns/leaf
+                              129,600 leaves/frame (exact: 3840·2160/64)
+                              (padded 64×64 accounting: 60×34 = 2,040
+                               CTUs/frame → 130,560 leaves at max split)
+Per-leaf budget: 16.67 ms / 129,600 = ~129 ns/leaf
+(Corrected 2026-07-16: the earlier 132,710 figure — 130,560 plus an
+unsourced "~1.6 % chroma alignment bias" — was not numerically
+grounded; use exact 129,600 or padded 130,560.)
 ```
 
 **Encoder per-leaf breakdown (scalar reference, current):**
@@ -816,11 +818,11 @@ Per-leaf budget: 16.67 ms / 132,710 = 125 ns/leaf
 | rANS encode (A7) | ~40 ns | ~40 ns |
 | **Total per-leaf** | **~960 ns** | **~210 ns** |
 
-**At scalar reference (960 ns/leaf): 4K @ 60 fps requires 132,710 ×
-960 ns = 127 ms/frame. Misses 60 fps by 7.6×.**
+**At scalar reference (960 ns/leaf): 4K @ 60 fps requires 129,600 ×
+960 ns = 124 ms/frame. Misses 60 fps by ~7.5×.**
 
-**At SIMD-batched (210 ns/leaf): 132,710 × 210 ns = 28 ms/frame. Misses
-60 fps by 1.7×; needs further work but in the same order of magnitude.**
+**At SIMD-batched (210 ns/leaf): 129,600 × 210 ns = 27 ms/frame. Misses
+60 fps by ~1.6×; needs further work but in the same order of magnitude.**
 
 **To hit 60 fps 4K real-time** requires the SIMD-batched-encode path
 to land. **This pins B:D-CODEC-8 / A:T-7 from P2 to P1.** Plan A4-impl
@@ -1261,7 +1263,7 @@ that decides whether each holy-grail claim is demonstrated.
 | R-8 (Plan G is confidence gate) | this doc | Phase order | Plan G ships before A7 | A7 PR doesn't merge until Plan G binary green |
 | R-9 (topology-free) | this doc | grep audit | Codec body has no spatial-semantic refs | `grep -rE 'North\|East\|West\|South' src/hpc/codec/*.rs` returns only test/doc |
 | R-10 (4 bit/Gaussian floor) | this doc | Plan E bench | bits/Gaussian on Mip-NeRF 360 | ≤4 bits/Gaussian without offline codebook |
-| R-11 (4K 60fps SIMD-batched) | this doc | Plan G video latency assert | Per-CTU encode time | ≤210 ns/CTU on Sapphire Rapids |
+| R-11 (4K 60fps SIMD-batched) | this doc | Plan G video latency assert | Per-leaf encode time | ≤210 ns/leaf on Sapphire Rapids (≡ ≤13.4 µs per fully-split 64×64 CTU = 64 leaves; corrected 2026-07-16 — the earlier "≤210 ns/CTU" mislabeled the per-leaf breakdown total) |
 | R-12 (per-CTU flush) | this doc | A8 frame-header parse + decode | First-CTU latency | First CTU decodable before frame complete |
 | R-13 (Option A per-shard) | this doc | Plan F on BERT-glue | 8× compression + accuracy | Holds; else escalate to Option B |
 | R-14 (Pillar 10 active) | this doc | `cargo test -p jc` (default features) | Pflug-Pichler Lipschitz bound | Pillar 10 probe green |
@@ -1414,7 +1416,7 @@ When this doc is summarised across context windows, preserve at minimum:
    - R-8: Plan G is confidence gate
    - R-9: topology-FREE codec layer
    - R-10: ~4 bits/Gaussian near target, ~1 bit stretch
-   - R-11: 210 ns/CTU SIMD-batched encode
+   - R-11: 210 ns/leaf SIMD-batched encode (per-leaf, not per-CTU)
    - R-12: per-CTU flush default; per-bucket Plan F
    - R-13: Option A (per-shard codebook) for Plan F v1; primitives are
      `cam_pq` + `bgz-hhtl-d` + `dn_tree` + `merkle_tree`

.claude/knowledge/pr-x12-woa-multiarch-orchestration.md

@@ -10,8 +10,7 @@
 > doc as evidence for anything.** Pending rewrite per audit Tier-1
 > recommendation; the W1a consumer contract at
 > `vertical-simd-consumer-contract.md` is the authoritative source.
-
-
+>
 > Date: 2026-05-22
 > Status: **perspective doc** — examines how the orchestration crates (`woa-rs`, `woa`, `q2`, `surrealdb`, `MedCare-rs`, `smb-office-rs`) consume the PR-X12 substrate, and how PR-X12's per-arch polyfill decisions (R-4, R-5, R-11) generalise to the entire HPC stack.
 >

.claude/knowledge/pr-x12-x265-blasgraph-gemm.md

@@ -104,7 +104,7 @@ Tropical-GEMM:    O(d²) × |nodes| = 16 × 85 = ~1.4 K ops/CTU
                   ~16× speedup
 ```
 
-For 4K @ 60 fps with 132K CTUs/frame, this is the difference between **4 ms and 64 ms per frame just for partition RDO**. At 60 fps's 16.67 ms budget, naive RDO doesn't fit.
+For 4K @ 60 fps with ~2,040 CTUs/frame (corrected 2026-07-16 — the earlier "132K CTUs" was the 8×8 leaf count), this is the difference between **~2.8 ms and ~44 ms per frame just for partition RDO** at ~1 op/ns. At 60 fps's 16.67 ms budget, naive RDO doesn't fit.
 
 **Dep direction:** the tropical-GEMM kernel's canonical home is `lance-graph::blasgraph` (the bit-exact cognitive-side substrate). *(Corrected 2026-07-16, audit #3: blasgraph's shipped semirings are binary-Hamming over 16384-bit BitVec — the numerical f32 min-plus kernel is UNWRITTEN and lands there when A6 wires it; the only shipped min-plus today is `bgz17::ScalarCsr::spmv_min_plus`, a lossy-sibling prototype.)* Post-Plan-H, `ndarray-codec → lance-graph::blasgraph` is allowed because both are sibling crates above `ndarray` hardware.
 
@@ -218,7 +218,7 @@ pub fn kmeans_predict_batched(
 );
 ```
 
-**Codec layer:** ~30-50 LoC per stage to wrap the BLAS call into the predict/A6/A4 flow. **BLAS layer:** zero new lines — all four already exist or land via existing infrastructure (`bf16_tile_gemm`, `cam_pq`, `simd_int_ops`).
+**Codec layer:** ~30-50 LoC per stage to wrap the BLAS call into the predict/A6/A4 flow. **BLAS layer:** the four wrapper APIs above still require implementation — what exists today is the underlying infrastructure they compose (`bf16_tile_gemm`, `cam_pq::kmeans` + `CamCodebook::distance_batch`, `simd_int_ops`); the wrappers are thin but not written (corrected 2026-07-16 — the earlier "zero new lines, all four already exist" overstated).
 
 This is what makes R-3's ≤1500 generic-codec-LoC ceiling reachable. Most of the heavy lifting is already in `blas_level2`; the codec adds wrappers and orchestration, not new BLAS code.