board: record the PR #258 arc — neutralize-by-cfg beats remove

Two appends, no code:

- EPIPHANIES.md: the #258 finding. The "dalek must be removed" conclusion
  was reached twice from a correct intrinsic count (57 sites) and a wrong
  inference — the whole vector backend sits behind one cfg at
  backend/mod.rs:42, so reachability, not porting, was the question.
  Records the verified counts, the one-line fix, the serial-costs-nothing
  argument, and the X25519 contributory-check tripwire.

- vertical-simd-consumer-contract.md: the escalation ladder this mints
  for third-party deps — (1) grep for the backend gate and neutralize by
  cfg, (2) port onto ndarray::simd only for paths we execute, (3) remove
  only with a failed gate-grep cited. "Contains raw intrinsics" and
  "raw intrinsics are reachable" are different claims; the matryoshka
  invariant binds the second.

Author: claude

.claude/board/EPIPHANIES.md

@@ -124,3 +124,51 @@ inspiration, never authority.
 fabrications still on master) · #5 ASG canon spec · #7 ASG-leaf impl (extend
 `CamPlaneSplat`, don't reinvent) · `cam-pq-production-wiring` (cam_pq shipped, unrouted
 through `CamCodecContract`) · `UNUSED_INVENTORY_1.95` A1–A9 dead-code.
+
+## 2026-07-28 — A dependency's SIMD surface is a reachability question, not a porting one (PR #258)
+**Status:** FINDING (verified — counts + cfg gate + build output all checked live)
+**Scope:** @simd-savant @workspace-primer domain:crypto domain:matryoshka
+
+The same wrong conclusion was reached twice in one arc: "curve25519-dalek's
+vector backend carries raw intrinsics (`avx2/field.rs` reaches around
+`packed_simd.rs`), therefore the matryoshka pattern requires removing the
+dependency." It got as far as a merged removal commit (`3694a3c`) before the
+premise was checked. The counts were right; the conclusion was not:
+
+- **Counts (verified):** 57 `_mm*` intrinsic calls in the AVX2 path
+  (35 `backend/vector/avx2/field.rs` + 22 `backend/vector/packed_simd.rs`),
+  under 52 `unsafe` occurrences in those two files. (`ifma/field.rs` adds 6
+  more on a nightly-only path.)
+- **The gate (verified):** ALL of it sits behind ONE cfg —
+  `curve25519-dalek-4.1.3/src/backend/mod.rs:42`,
+  `#[cfg(curve25519_dalek_backend = "simd")] pub mod vector;` — and dalek's
+  build.rs reads that cfg from `CARGO_CFG_CURVE25519_DALEK_BACKEND`.
+- **The fix (one line):** `.cargo/config.toml` rustflags
+  `--cfg curve25519_dalek_backend="serial"` compiles the entire second SIMD
+  surface out. Verified after `cargo clean -p curve25519-dalek`: the build
+  emits `curve25519_dalek_backend="serial"` only.
+- **`serial` costs nothing we use:** X25519's Montgomery ladder
+  (`montgomery.rs`) never references the vector backend; the vector path
+  serves only Edwards multi-scalar / variable-base multiplication (batch
+  verification, Ristretto protocols) — neither performed by
+  `crates/encryption`.
+
+**The rule this mints:** before porting a third-party crate's SIMD onto
+`ndarray::simd` — and *long* before removing the crate — grep for the
+backend's cfg/feature gate. "It contains raw intrinsics" and "raw intrinsics
+reach the binary" are different claims; the matryoshka invariant binds the
+second, not the first. PR #258 (merged `b127e25`) is the worked example: the
+removal reverted, Ed25519 kept, the surface neutralized by cfg, diff net
+additive.
+
+**Carried tripwire (X25519 port):** `x448::x448()` returns `Option` and
+rejects low-order points; `x25519_dalek::x25519()` returns a bare
+`[u8; 32]` — RFC 7748's contributory check is OPTIONAL, so a mechanical port
+drops it silently. The test
+`channel::tests::low_order_peer_keys_are_refused_and_honest_ones_are_not`
+(both halves asserted, per the falsifiability rule) fails if it does.
+
+Cross-ref: PR #258 body (the full correction narrative), `.cargo/config.toml`
+comment block (the in-tree record), lance-graph
+`E-VACUOUS-ASSERTION-IS-THE-HOUSE-STYLE-1` (the falsifiability rule the
+low-order test follows).

.claude/knowledge/vertical-simd-consumer-contract.md

@@ -326,3 +326,32 @@ The `simd-savant` agent on the `lance-graph` side runs PRE-MERGE against every W
 
 > **Is the consumer site cited in the PR description?**
 > Missing = reject. We're shipping primitives for known workloads, not speculative ones.
+
+## Third-party dependencies: neutralize by cfg BEFORE porting, port BEFORE removing (2026-07-28, PR #258)
+
+The matryoshka invariant — all SIMD lives once, audited, inside
+`ndarray::simd` — binds what **reaches the binary**, not what a dependency's
+source tree *contains*. When a third-party crate carries its own intrinsics,
+the escalation ladder is:
+
+1. **Grep for the gate.** Most crypto/perf crates gate their vector backend
+   behind one cfg or feature (`curve25519-dalek`:
+   `#[cfg(curve25519_dalek_backend = "simd")] pub mod vector;` at
+   `backend/mod.rs:42`, selected by build.rs from
+   `CARGO_CFG_CURVE25519_DALEK_BACKEND`). One rustflags line in
+   `.cargo/config.toml` compiles the whole surface out. Cost: zero code.
+   Verify with a clean rebuild that the cfg actually flips
+   (`cargo build -v | grep <cfg name>`).
+2. **Port onto `ndarray::simd`** only if the intrinsics are load-bearing for
+   a path we actually execute (the `vendor/chacha20` precedent — the ARX
+   lane IS our hot path).
+3. **Remove the dependency** only if neither applies — and then the claim
+   "can't be neutralized" must cite the failed gate-grep, not the intrinsic
+   count.
+
+PR #258 is the cautionary worked example: an intrinsic count (57 sites, real)
+was read as a porting obligation, escalated straight to step 3, and a merged
+removal commit had to be reverted once the step-1 gate was found. "Contains
+raw intrinsics" and "raw intrinsics are reachable" are different claims;
+audit the second. Full record: board `EPIPHANIES.md` 2026-07-28 entry +
+the `.cargo/config.toml` comment block.