Summary

• Adds a content-plane Hamming pre-pass to ShaderDriver::run() before the existing palette cascade. For every pair in passed_rows, popcount the XOR of their content fingerprints; if resonance = 1 - Hamming/16384 >= style.resonance_threshold, emit a ShaderHit{predicates: 0x01} (content-match flag).
• Result: dispatch() now surfaces real text similarity. Previously the PaletteSemiring cascade probed a synthetic Base17 distance table unrelated to the encoded text, and the content plane was read only for the cycle_fp XOR fold, never compared. The audit in EPIPHANIES 2026-04-24 "Ground truth: ShaderDriver dispatch wiring audit" called out Option A (content Hamming pre-pass) as the fix — this PR wires it.
• Palette cascade kept intact (complementary edge-palette signal). Guard: skip the N^2 sweep when passed_rows.len() > 256.

The problem, before

encode "Palantir develops surveillance systems"     -> row 0 (247 bits)
encode "Palantir Gotham is a surveillance platform" -> row 1 (249 bits)
encode "Israel deploys military AI"                 -> row 2 (259 bits)
dispatch rows 0..3, style=analytical
  -> hit_count: 0, top_k: [], confidence: 0.0, admit_ignorance: true

Zero hits across every style. The Jirak signal (3σ at Hamming < 454) was present in the data, but invisible to dispatch().

The fix, after

dispatch rows 0..3, style=creative (threshold 0.35)
  hit_count: 6
  top_k:
    row 0, distance 6592, predicates 0x01, resonance 0.598   <- Palantir / Palantir pair
    row 1, distance 6592, predicates 0x01, resonance 0.598
    row 1, distance 7232, predicates 0x01, resonance 0.559
    row 2, distance 7232, predicates 0x01, resonance 0.559
    row 0, distance 8256, predicates 0x01, resonance 0.496
    row 2, distance 8256, predicates 0x01, resonance 0.496
  confidence 0.598, admit_ignorance: false

The strongest pair (rows 0↔1, both Palantir) correctly ranks first. Rows 0↔2 (Palantir vs Israel AI) lands lowest. Analytical's 0.85 threshold correctly rejects all pairs (none cross 0.85). Peripheral (0.20) matches creative here — same pairs survive because the lowest pair is still at 0.496 > 0.20.

Threshold model

Resonance-over-threshold rather than an absolute bit count, because the 32x DeepNSM tiling pushes content-plane density from 0.016 to ~0.48 and the Jirak-calibrated 454-bit reference no longer applies directly. Per-style thresholds from UNIFIED_STYLES.resonance_threshold:

Cross-ref: EPIPHANIES 2026-04-24 "Jirak noise floor calibrated for DeepNSM-tiled 16K-bit fingerprints" + "Ground truth: ShaderDriver dispatch wiring audit".

Test plan

• cargo test -p cognitive-shader-driver — 45 pass (43 lib + 2 e2e, 3 new)
  • content_hamming_finds_similar_rows — two rows differ in 4 bits, analytical finds them at resonance > 0.5.
  • content_hamming_skips_dissimilar — two rows with Hamming ~10000 (resonance ~0.39), analytical correctly emits 0 content hits.
  • content_hamming_respects_style_threshold — Hamming ~5000 (resonance ~0.695), analytical 0 content hits, creative >= 1 content hits; monotonicity assertion.
• Live verification: run shader-serve, encode 3 related rows, dispatch analytical/creative/peripheral; observe predicates=0x01 hits with correct pair ordering.
• No regressions in pre-existing dispatch tests (meta prefilter, sink short-circuit, cycle fingerprint emission).

Known follow-ups

• Palette cascade hits still intermix via shared resonance sort — in production with meaningful planes, high-resonance palette hits can truncate content-match hits out of top_k (observed in the respects-style-threshold test when using demo_planes; resolved there by isolating with empty planes). If future tuning shows content drowning, consider a small content-hit resonance bonus or preserving the top-K content + top-K palette separately before the final merge.
• The Jirak-derived absolute Hamming threshold (454 at 3σ) is density-dependent; document on the encode handler that the 32x tiling deviates from the calibration baseline. Not a blocker for this PR.

https://claude.ai/code/session_01SbYsmmbPf9YQuYbHZN52Zh

Generated by Claude Code