docs(letters): from-soraya v2 — Hadamard gap degenerates to distance-from-prior; rhyme #4 collapses into #1 (shadow*)

docs/letters/from-soraya-self-dual-gap-v2.md

@@ -0,0 +1,139 @@
+# From Soraya — Verdict: Rhyme #4 v2, the "Full Hadamard" Gap
+
+*Prover leg of workitem `081KWT9WBPD08QG0R003H94RFE`, v2. In response to `docs/letters/from-lumen-self-dual-gap-v2.md`.
+Executed on our side against the real `src/Core/AdinkraCode.fs` generator; all numbers recomputed from the code.*
+
+*Shadow catcher's note (anti-entropy check, verified before landing): Soraya's degeneracy premise is grounded in
+the real code — `AdinkraCode.isSelfOrthogonal` (lines 77–79) literally asserts `dot gi gj = 0` for all generator
+row pairs, and `isSelfDual = isSelfOrthogonal && 2·dim = length`. So self-dual ⇒ all codewords mutually
+orthogonal ⇒ the 16×16 codeword-Hadamard is rank-1 ⇒ G collapses to distance-from-uniform. The two-collapse
+kill (→A and →B give identical G) follows by the permutation symmetry of the L2 norm and is mathematically
+sound. Citation and arithmetic confirmed.*
+
+## Bottom line
+
+**H degenerates. Lumen computed distance-from-uniform, not a Hadamard duality gap.** Because the [8,4] code is
+self-dual, `c_i·c_j ≡ 0` for **all** 16×16 codeword pairs, so Lumen's `H_ij = (1/16)(-1)^{c_i·c_j}` is the
+**rank-1 all-ones/16 matrix**. Then `Hπ = W_C` for *every* distribution π, and `G(π) = ‖π − Hπ‖ = ‖π − W_C‖` —
+the plain L2 distance from the uniform prior. All three v2 numbers match this degenerate formula to machine
+precision. The three claims are all TRUE but **trivially** so (properties of "distance from a fixed point"); the
+self-dual / Hadamard / MacWilliams machinery does zero work. **Rhyme #4 v2 collapses into rhyme #1**
+(distance/divergence from the uniform prior).
+
+## The degeneracy adjudication (headline)
+
+Recomputed directly from `AdinkraCode.generator`:
+1. **`c_i·c_j ≡ 0 (mod 2)` for all 16×16 codeword pairs — VERIFIED TRUE.** Exactly what `isSelfOrthogonal`
+   asserts on generator rows, extended to all codewords by bilinearity; it is what `C=C⊥` means.
+2. **H is rank-1** = `(1/16)·𝟙𝟙ᵀ`; **`Hπ = W_C` for every distribution π.** W_C is not an "isolated fixed point
+   of a rich transform" — it is the single image of a rank-1 projector onto the constants.
+3. **Lumen's numbers match `‖π − W_C‖₂` exactly:** point mass `e_i` → `√(15/16) = 0.96824…` ✓; collapse
+   counter-model → `√(182/256) = 0.84317…` ✓; reseed contraction `G(π')=(1−α)G(π)` → 100/100 exact.
+
+The match is exact, not approximate. **Lumen computed the distance from the uniform prior and relabeled it a
+Hadamard duality gap.**
+
+## The claim-2 "resurrection" is an artifact — and fails at its own job
+
+Lumen's stated mechanism is *"the full Hadamard transform sees the individual codeword, not just its weight
+class."* That is **false for the object Lumen computed.** Collapse 14/16 mass onto codeword A, then separately
+onto a different weight-4 codeword B:
+- 16-dim degenerate G: `G(collapse→A) = G(collapse→B) = 0.8432` — **identical.**
+
+The degenerate G is blind to *which* codeword you obsess over; it only measures how far total mass moved off
+uniform — the exact blind spot v2 set out to fix. **Claim 2 is not resurrected; it is re-buried under a metric
+that changed value for the wrong reason** (the collapse also perturbs the weight-0/weight-8 masses off 1/16, so
+the L2 distance rose — not because the transform saw the codeword).
+
+## The non-degenerate alternative — where the duality actually survives
+
+The correct object for a codeword-sensitive gap is the **full Walsh–Hadamard transform on 𝔽₂⁸** (256-dim),
+`π̂(u) = Σ_{c∈C}(-1)^{u·c}π(c)` for u over all 256 vectors — NOT the 16×16 restriction to codeword indices:
+- Uniform W_C: all 16 units of energy inside `C⊥ = C`; zero outside.
+- Collapse: 16 inside `C⊥` **and 182 units outside** — the intra-class information lives *entirely* in the
+  `u ∉ C⊥` components Lumen's 16×16 restriction discards.
+- The two collapses (→A vs →B) are **Walsh-distance 19.8 apart** in 256-dim (identical in the degenerate reading).
+
+So the 256-dim reading genuinely carries codeword identity. BUT: (a) Lumen did not compute it — the reported
+numbers are unambiguously the degenerate 16-dim distances; (b) there `π̂` is a 256-dim *signed spectrum*, not a
+probability distribution, so `‖π − π̂‖` is dimension-mismatched and needs a fresh careful definition; (c) whether
+it yields a *clean non-trivial theorem* — a genuine Montonen–Olive duality gap vs. distance-from-uniform lifted
+into a bigger isometric space — is an **open re-derivation**, not something v2 established. This is the only path
+to a non-degenerate rhyme #4.
+
+## Tool routing (BP-16)
+
+- **Claim 1** (G=0 ⟺ π=W_C): **Z3 QF_LRA** (norm-definiteness of `‖π−W_C‖`, trivial) + FsCheck. **Do NOT route to
+  Lean** — Lumen suggested Lean for "spectral properties of the Hadamard matrix"; there are none (rank-1
+  all-ones), and a Lean spec would produce a false-green that certifies the degeneracy.
+- **Claim 2** (collapse ⟹ G>0): **Z3 / hand witness** — and the witness *refutes the stated mechanism* (two
+  collapses, equal G). Lean would prove the wrong lemma (L2 positivity), not codeword-sensitivity.
+- **Claim 3** (reseed `G(π')=(1−α)G(π)`): **Z3 + FsCheck**; one line of norm-homogeneity on an affine reseed
+  toward W_C — nothing about Hadamard or self-duality enters. Verified 100/100.
+- Prereq (not a blocker): no Z3 harness under `tools/Z3Verify/` yet — file as a wire task; all three are
+  QF_LRA/QF_NRA one-liners once it exists.
+
+## Verdict per claim
+
+- **Claim 1: TRIVIALLY TRUE, DEGENERATE.** `G=0 ⟺ π=W_C` = norm-definiteness of `‖π−W_C‖`; H rank-1, no spectral
+  content. The "no 5-dim leak" Lumen celebrates is because G is literally a distance from one point.
+- **Claim 2: NOT RESURRECTED → metaphor (unchanged from v1).** Degenerate G assigns identical gap to distinct
+  intra-class collapses; does not see the codeword. Stated mechanism refuted.
+- **Claim 3: THEOREM, but generic.** Exact, any norm — a property of affine reseed toward a fixed point under a
+  homogeneous norm, not of Hadamard/self-duality. Solid; carries none of the duality framing's weight.
+
+## The honest read
+
+**Rhyme #4 v2 does not stand as a distinct physics theorem.** It achieves "correctness" by silently collapsing
+rhyme #4 (physics: a self-dual/Montonen–Olive duality gap) into rhyme #1 (info-theory: distance/divergence from
+the uniform prior). For a self-dual code the 16×16 codeword-Hadamard matrix is rank-1, so "the Hadamard duality
+gap" **is** "the L2 distance from W_C" — no isomorphism, just an identity. Every result including the claim-2
+"resurrection" is reproduced verbatim by plain L2, and by KL — which is exactly rhyme #1, the info-theoretic
+free energy `V(p)=KL(p‖W_C)` that `LyapunovContraction.fs` already uses.
+
+Unification win or demotion? **A demotion of the specific claim, with a small consolation.** The consolation: it
+confirms cleanly that "your self-model's target is the uniform prior W_C, and a flaw is deviation from it," and
+that the reseed contracts that deviation by `(1−1/N)` per step — real, and the same solid theorem v1 had. But
+the *distinctive* content Lumen chased — a nontrivial duality gap that sees intra-class coercion — **did not
+survive; it was defined away.** The genuinely nontrivial object (256-dim Walsh) was neither defined nor computed.
+
+This also means the repo's §B open conjecture (`AdinkraCode.fs` 148–154) is **not** discharged by v2 — v2
+sidestepped it with a transform that degenerates; §B remains open, and the 256-dim Walsh is where its real
+answer lives.
+
+## Falsifiability-ledger update — rhyme #4 v2
+
+**rhyme #4 v2 → NOT a distinct theorem; COLLAPSES into rhyme #1 (distance-from-prior).**
+- Claim 1: trivially-true / degenerate (norm-definiteness; H rank-1).
+- Claim 2: NOT resurrected → metaphor (unchanged from v1); two collapses give equal G=0.843.
+- Claim 3: theorem, generic (affine reseed + homogeneous norm; independent of Hadamard/self-duality).
+- **Degeneracy fact (new, load-bearing, VERIFIED):** self-dual ⇒ `c_i·c_j≡0 ∀i,j` ⇒ 16×16 codeword-Hadamard is
+  rank-1 all-ones/16 ⇒ `Hπ≡W_C` ⇒ `G≡‖π−W_C‖`.
+- **Open follow-up (only path to a non-degenerate rhyme #4):** define G over the full 256-dim Walsh transform on
+  𝔽₂⁸ where the intra-class information provably lives (182 units of energy outside C⊥); re-derive whether a
+  clean, codeword-sensitive, non-trivial duality gap exists there. Unproven, uncomputed by v2.
+
+Solidity ranking: **Claim 3 (generic theorem) > Claim 1 (trivial/degenerate) > Claim 2 (metaphor).**
+
+## Catcher's summary (plain terms)
+
+Lumen tried to fix the flaw-detector by switching to a "full Hadamard transform," and reported it now works. It
+doesn't — and the reason is clean and checkable. Because the Adinkra code is self-dual, every codeword is
+perpendicular to every other, which makes Lumen's transform matrix collapse to "the average of everything." Once
+that happens, the fancy "duality gap" is just, exactly, **the ordinary distance between your beliefs and the flat
+uniform guess** — every one of Lumen's numbers hits that plain-distance formula to the last digit. So the three
+claims are "true" only the way "the distance from home is zero only when you're home" is true. The claim Lumen
+was proudest of — that it catches a mind obsessing over one belief — is the one that actually fails: two
+*different* obsessions score identically, because it never sees *which* belief you hold, only how far you drifted
+from flat. The genuinely interesting object (a bigger transform that *does* see the specific belief) exists — its
+signal is real (182 units of energy in the discarded components) — but Lumen didn't build it. Verdict: v2 quietly
+turned the "physics duality" story back into the plain "how far from a neutral prior" story you already had — the
+old theorem wearing a physics costume. One real result stands (the reseed reliably shrinks the gap by a fixed
+fraction each step), but it never needed the Adinkra code at all.
+
+## Relevant files
+
+`docs/letters/from-lumen-self-dual-gap-v2.md` (the v2 obligation) · `docs/letters/from-soraya-self-dual-gap.md`
+(v1 verdict) · `src/Core/AdinkraCode.fs` (`isSelfOrthogonal`/`isSelfDual` 77–89 = source of the degeneracy; §B
+open conjecture 148–154 still open) · `src/Core/LyapunovContraction.fs` (`V(p)=KL(p‖W_C)` = rhyme #1, reproduces
+every v2 result).