finding: ground truth audit — what dispatch() actually wires vs promises

Honest audit of ShaderDriver.dispatch() (driver.rs:75-212):
- CASCADE probes PaletteSemiring on edge palette indices, NOT content
  fingerprints. Content plane is read (XOR fold) but never compared.
- NARS types exist but aren't dispatched — style_ord_to_inference()
  packs into CausalEdge64 but no actual NARS runs.
- FreeEnergy type exists (grammar/free_energy.rs) but dispatch()
  never calls compose(). admit_ignorance is threshold, not F.
- AriGraph/SPO not connected to driver. PropertySchema not connected.

Three options to wire content similarity into the cascade documented.
This is the gap between "the glove fits" and "the glove flies."

https://claude.ai/code/session_01SbYsmmbPf9YQuYbHZN52Zh

Author: claude

.claude/board/EPIPHANIES.md

@@ -2783,3 +2783,54 @@ Grounding the NaN: with DeepNSM encode (512-bit VSA tiled 32× into 16K), densit
 **Jirak citation:** Jirak 2016, arxiv 1606.01617, Annals of Probability 44(3). Rate: n^(p/2-1) for p in (2,3]. Weak dependence sources: (a) tiling (32x repeat of 512-bit), (b) XOR-bind braiding, (c) FNV-1a hash collision at 12-bit rank.
 
 Cross-ref: I-NOISE-FLOOR-JIRAK iron rule, encode_handler, DeepNSM VsaVec::from_rank().
+
+## 2026-04-24 — Ground truth: ShaderDriver dispatch wiring audit (what IS vs ISN'T connected)
+
+**Status:** FINDING
+**Owner scope:** @truth-architect, @bus-compiler
+
+Honest audit of what dispatch() actually does vs what the DTO surface promises:
+
+**WIRED (working end-to-end):**
+- [1] Meta prefilter: u32 column sweep on MetaColumn → passed_rows ✓
+- [2] Style resolution: Auto reads QualiaColumn of first row → style_ord ✓
+- [3] Shader cascade: CognitiveShader::new(planes, semiring).cascade(query, radius, layer_mask) ✓
+  BUT: query comes from CausalEdge64.s_idx() of the ROW'S EDGE, not from content fingerprint.
+  The cascade probes the PaletteSemiring distance table, not the content plane.
+- [4] Cycle fingerprint: XOR fold of content_row(hit.row) for each hit ✓
+  BUT: hits come from step [3] which probes edges, not content similarity.
+- [5] Entropy + std_dev + CollapseGate: computed from top-k resonances ✓
+- [6] Edge emission: CausalEdge64::pack per strong hit ✓
+- [7] Sink callbacks: on_resonance → on_bus → on_crystal ✓
+- Meta summary: confidence = top-1 resonance, admit_ignorance = confidence < 0.2 ✓
+
+**NOT WIRED (the gap):**
+- Content fingerprint similarity: dispatch does NOT compare content_row(A) vs content_row(B).
+  The cascade uses PaletteSemiring on edge palette indices, not Hamming on content bits.
+  The content plane is READ (for cycle_fp XOR fold) but never COMPARED.
+- NARS reasoning: no InferenceType dispatch. style_ord maps to inference type via
+  style_ord_to_inference() but it's only used for CausalEdge64 packing, not actual NARS.
+- FreeEnergy: not computed. The contract type exists (grammar/free_energy.rs) but
+  dispatch() never calls FreeEnergy::compose(). The 'should_admit_ignorance' is a
+  simple threshold (confidence < 0.2), not a real F computation.
+- AriGraph/SPO: no graph. dispatch() operates purely on BindSpace columns.
+  The SPO triple store exists in lance-graph core but isn't wired to the driver.
+- PropertySchema validation: not connected. The types exist in contract::property
+  but dispatch() doesn't check Required/Optional/Free.
+
+**What the zeros meant:** resonance=0 wasn't "missing semiring wire" — the cascade
+DID run (3 cascade calls from step [3]). But the demo palette has synthetic Base17
+entries with no relationship to the encoded text. The PaletteSemiring distance table
+is 256x256 pre-computed from those synthetic entries. Text fingerprints in the content
+plane are INVISIBLE to the cascade — they're read only for the XOR fold in step [4].
+
+**To make content fingerprints visible to dispatch:**
+Option A: Add a HammingMin pre-pass before the palette cascade. Compare content_row(i) vs
+  content_row(j) via popcount on XOR. If Hamming < Jirak threshold (454), promote to hit.
+Option B: Build the PaletteSemiring FROM the content fingerprints (quantize content into
+  256 palette entries, compute distance table from those). Content similarity then flows
+  through the existing cascade.
+Option C: Add a second dispatch mode (content-mode vs edge-mode) that uses HammingMin
+  instead of PaletteSemiring for the distance function.
+
+Cross-ref: driver.rs:75-212, Jirak calibration (this session), I-NOISE-FLOOR-JIRAK.