docs(research): attack plan — from suspect to receipt

Every result is suspect until receipt matches conjecture. 10 conjectures
(C1-C10) with predictions, receipts, and rejection criteria.

Phase 0: VERIFY FOUNDATION (blocking)
  C1: CRTSidonN compiles? → lake build (running)
  C2: HCMR suite compiles? → same build
  C7: HCMR self-loops measured? → search Research Stack

Phase 1: TEST THE PIPELINE (blocking)
  C3: Positional chirality varies? → run pipeline_core.py
  C4: Quaternion filter varies? → run pipeline_core.py --filter quat
  C5: Kelvin check filters? → count Kelvin-rejected

Phase 2: GPU (after Phase 1)
  C6: Cross-enrich shader runs? → compile on GPU
  C9: de Grey Hoffman bound? → run --full

Phase 3: ROBUSTNESS
  C8: q-profile robust? → 10+ label sets
  C10: Rossby ↔ QUBO? → correlation experiment

Rule: no conjecture accepted until receipt matches prediction.
Each failure narrows the theory to what's actually true.

docs/research/ATTACK_PLAN.md

@@ -0,0 +1,153 @@
+# Attack Plan: From Suspect to Receipt
+
+**Status:** ACTIVE — every result suspect until receipt matches conjecture
+**Date:** 2026-07-04
+**Principle:** No claim is accepted without a receipt. Conjecture → prediction → measurement → receipt → accept or reject.
+
+---
+
+## The 13 Things We Actually Have
+
+| # | What | Type | Receipt |
+|---|------|------|---------|
+| 1 | sidon_preserved_mod (2-moduli CRT Sidon) | Lean proof | CRTSidon.lean, 0 sorries, lake-built (verified on prior runs) |
+| 2 | helical_coverage_74 (74 steps → 28 classes) | Lean proof | HopfFibration.lean, native_decide |
+| 3 | ofChiralLabel_isUnit (ChiralLabel → unit quat) | Lean proof | HopfFibration.lean |
+| 4 | ring_fastest (ring > SUBLEQ > AVX-512) | Lean proof | HCMR.lean, simple omega |
+| 5 | Conservation law (program + residual ≥ K(data)) | Measured | 8 experiments, real bytes, weird_machine_conservation_law.md |
+| 6 | Chiral invariance (flat CRT negation, odd L) | Proven + measured | 50K trials, ring automorphism proof |
+| 7 | Hoffman gap=1 for unit-distance graphs | Measured | numpy eigenvalues, 6 graphs, hn_spectral_database.json |
+| 8 | q-profile sweep (q>1 = 100% Sidon) | Measured | crt_qprofile_sweep.json, exact arithmetic |
+| 9 | Photonic Sidon 18/18 PASS | Measured | Perceval SLOS, photonic_sidon_evidence.jsonl |
+| 10 | pipeline_core.py runs (binary swaps) | Implemented | chiral_batch_pipeline.json, 256→64 |
+| 11 | dna_braid.wgsl (GPU braid sort) | Implemented | Existing, tested |
+| 12 | dna_radix_gpu.py (GPU QUBO sort) | Implemented | Existing, tested |
+| 13 | BraidStateN.lean ChiralLabel/Rossby | Lean, compiles | Prior lake build passed |
+
+---
+
+## Conjectures to Validate (in priority order)
+
+### C1: CRTSidonN.lean compiles
+**Conjecture:** The n-moduli generalization is valid Lean that compiles.
+**Prediction:** `lake build SilverSight.CRTSidonN` exits 0.
+**Receipt needed:** Build log showing 0 errors.
+**Status:** Lake build running (currently at ~73% of full Mathlib).
+**Action:** Wait for build, check exit code.
+
+### C2: HCMR suite compiles
+**Conjecture:** The 5 new modules (HCMR, CacheSieve, Blitter6502OISC, YangMillsPerformance, WorkloadTestbench) compile with 2 sorries.
+**Prediction:** `lake build` exits 0 with exactly 2 sorries (CacheSieve.evict_prefers_reset, YangMillsPerformance.compression_overhead_bounded).
+**Receipt needed:** Build log + sorry count.
+**Status:** Same lake build run.
+**Action:** Check build output for error count and sorry locations.
+
+### C3: Positional chiral pipeline produces non-uniform Sidon results
+**Conjecture:** The positional chirality (permuting phases across strand positions) produces DIFFERENT Sidon results for different chiral configs.
+**Prediction:** Not all 2^k configs have the same collision count (unlike flat negation, which was 100% uniform).
+**Receipt needed:** pipeline_core.py output with `--filter crt` showing variance in collisions across configs.
+**Status:** UNTESTED. The previous run used flat negation (chiral-invariant).
+**Action:** Run pipeline_core.py with positional chirality on CPU compute. Compare collision counts across configs. If all identical → positional chirality is also invariant (conjecture FALSE). If variance exists → conjecture TRUE.
+
+### C4: Quaternion product Sidon filter discriminates chiral configs
+**Conjecture:** The QuaternionSidonFilter (Hamilton product of 1,i,j,k basis) produces different results for different chiral configs.
+**Prediction:** Some configs pass (0 collisions), others fail (>0 collisions).
+**Receipt needed:** pipeline_core.py output with `--filter quat` showing variance.
+**Status:** UNTESTED. Previous DQ run gave 0/64 (all fail) — but that was with flat negation, not positional.
+**Action:** Run with positional chirality + quaternion filter. If variance → TRUE. If uniform → FALSE.
+
+### C5: COUCH Kelvin check actually filters
+**Conjecture:** Configs with Rossby drift = 0 (Kelvin regime) exist in the 65K cross-enriched space and are rejected by COUCH.
+**Prediction:** Some of the 65K configs have drift=0 (all-achiral or balanced left/right). These should be rejected.
+**Receipt needed:** COUCH pass count < total, with some configs tagged "kelvin".
+**Status:** UNTESTED.
+**Action:** Run pipeline with k=8, count Kelvin-rejected configs.
+
+### C6: Cross-enrichment (4^8 = 65K) runs on GPU
+**Conjecture:** chiral_cross_enrich.wgsl compiles and runs, processing 65K configs.
+**Prediction:** GPU dispatch completes, output buffer has 65K results.
+**Receipt needed:** WGSL compilation + dispatch log + output JSON.
+**Status:** UNTESTED. Shader is written but never compiled.
+**Action:** Test on GPU pod (A40 or user's WebGPU node). Check compilation, then run.
+
+### C7: HCMR self-loop probabilities are measured (not just defined)
+**Conjecture:** The values 0.823, 0.885, 0.0 come from real EPYC KVM benchmarks.
+**Prediction:** A benchmark log or measurement file exists with these values.
+**Receipt needed:** Benchmark output showing these numbers.
+**Status:** SUSPECT. The Lean module DEFINES them as constants. No measurement file found.
+**Action:** Search Research Stack for benchmark logs. If none found → relabel as "assumed" not "measured".
+
+### C8: q-profile sweep result (q>1 = 100% Sidon) is robust
+**Conjecture:** The q>1 preference holds across different label sets and moduli.
+**Prediction:** Reproducing with different Sidon label sets gives the same q>1 preference.
+**Receipt needed:** Re-run with 3+ different label sets, check q>1 rate.
+**Status:** Measured once (3 label sets). Need more.
+**Action:** Run q-profile sweep with 10+ random Sidon label sets.
+
+### C9: Hoffman gap=1 is universal for unit-distance graphs
+**Conjecture:** All unit-distance graphs have Hoffman gap=1 (not just Moser/Golomb).
+**Prediction:** de Grey 1581 graph also has gap=1.
+**Receipt needed:** Hoffman bound on de Grey 1581.
+**Status:** UNTESTED (the --full flag never ran successfully).
+**Action:** Run hn_spectral_database.py --full on GPU pod (needs numpy for 1581×1581 eigenvalues).
+
+### C10: Rossby drift correlates with QUBO tractability
+**Conjecture:** QUBO instances with Rossby drift=0 (Kelvin) are harder for QAOA.
+**Prediction:** Kelvin-regime QUBO instances have flatter energy landscapes.
+**Receipt needed:** QUBO energy landscape measurement + Rossby drift correlation.
+**Status:** ENTIRELY SPECULATIVE. No experiment designed.
+**Action:** Design experiment: generate random QUBO instances, compute Rossby drift from chiral encoding, measure energy landscape flatness, check correlation.
+
+---
+
+## Execution Order
+
+```
+Phase 0: VERIFY FOUNDATION (blocking)
+  C1: CRTSidonN compiles?        → lake build (running)
+  C2: HCMR suite compiles?       → same build
+  C7: HCMR self-loops measured?  → search Research Stack
+
+Phase 1: TEST THE PIPELINE (blocking)
+  C3: Positional chirality varies?  → run pipeline_core.py --filter crt
+  C4: Quaternion filter varies?     → run pipeline_core.py --filter quat
+  C5: Kelvin check filters?         → count Kelvin-rejected in C3/C4 output
+
+Phase 2: GPU (after Phase 1 confirms variance)
+  C6: Cross-enrich shader runs?     → compile + dispatch on GPU
+  C9: de Grey Hoffman bound?        → run --full on GPU
+
+Phase 3: ROBUSTNESS (after Phase 2)
+  C8: q-profile robust?             → 10+ label sets
+  C10: Rossby ↔ QUBO?               → design + run correlation experiment
+
+Phase 4: QUANTUM (after Phase 3 confirms C10)
+  QUBO/QAOA integration experiments
+```
+
+---
+
+## Rule
+
+Every conjecture has:
+1. A PREDICTION (what we expect)
+2. A RECEIPT needed (what would prove it)
+3. A REJECTION criterion (what would disprove it)
+
+No conjecture is accepted until the receipt matches the prediction.
+If the receipt contradicts the prediction → conjecture is DEAD, update the theory.
+
+If C3 fails (positional chirality is also invariant):
+  → The entire chiral filtering story is dead
+  → COUCH (geometric) is the only discriminating filter
+  → Sidon filter is decorative, not functional
+
+If C4 fails (quaternion filter is uniform):
+  → The quaternion/S³ story is dead
+  → Stick with CRT sums (proven, 2-moduli)
+
+If C10 fails (Rossby ≠ QUBO tractability):
+  → The QUBO/QAOA bridge is dead
+  → The pipeline is a combinatorial filter, not a quantum pre-filter
+
+Each failure narrows the theory to what's actually true.