sigker: delegate randomized-signature encode to the ndarray SIMD sweep

[AdaWorldAPI]

Summary

Closes the follow-up ndarray PR #294 explicitly deferred ("Wiring sigker to actually call this is out of scope").

Until now the primitive was dead code: ndarray shipped, tested and benchmarked the SIMD recurrence, but RandomizedSignatureBuilder::encode still ran its own row-major scalar loop — so the measured 2.0×–3.8× speedup was unrealised. encode now delegates to ndarray::hpc::randomized_signature::randomized_signature_sweep, mirroring the precedent kernel.rs:35 already set for PR #293's signature_pde_sweep.

The buffer layout was already identical (matrices[i*k*k + row*k + col], biases concatenated) and so is the |Δx_i| < 1e-15 skip, so the delegation carries sigker's numerics unchanged.

Behaviour change — deliberate, documented, tested

Both of sigker's stricter caller-facing asserts are kept ahead of the delegation:

assert!(!path.is_empty(), "path must have ≥1 point");
assert_eq!(path[0].len(), self.path_dim, "path point dim mismatch");

The ndarray primitive accepts an empty path and returns the zero state, so dropping these would have been a silent contract change.

But the primitive additionally asserts that every path point matches path[0]'s dimension. sigker previously checked only path[0], so a ragged path used to produce a silently truncated wrong signature and now panics. That is a strict improvement, recorded in a # Panics section and pinned by a test — not hidden.

Tests

Both new, both discriminating:

• encode_matches_scalar_reference_within_tolerance — parity against an inlined scalar oracle transcribed from the pre-delegation loop, at 1e-9 relative tolerance rather than bit-equality. The SIMD GEMV reduces eight partial sums and fuses products, and reduce_sum order differs per backend, so bit-equality is the wrong contract by construction. Catches drift in the recurrence, the skip epsilon, the activation, or the buffer layout.
• ragged_path_now_panics — the exact input that did not panic before this change and produced a wrong signature instead.

Board hygiene (same commit, per the Mandatory Board-Hygiene Rule)

EPIPHANIES — E-A-SKETCH-THAT-MISSED-TWICE-WILL-MISS-A-THIRD-TIME-1. ndarray's W1.5 lane-type sketches are 3-for-3 wrong:

item	doc sketch	real consumer	shipped as
#6 signature_pde	&[F32x16]	f64	F64x8
#7 randomized	F32x16 state	f64/Vec<f64>	F64x8
#8 lyndon-pack	I16x16 state	Vec<f64> (log_signature.rs:272)	—

The entry also records a stronger finding: W1.5-#8 is not a well-shaped SIMD primitive at all. "Pack/unpack" names no operation that exists; the real cost centres are bracket_expansion (:227-256, recursive sparse outer-product with sort/coalesce) and a data-dependent scatter-add peel (:368-386) — the opposite of a dense fixed-lane kernel. The basis is generated on the fly by Duval's algorithm (:164, :346) with no table to consume. The actual bottleneck the module names itself (:100) is the Magnus expansion in tensor_multiply, a different primitive. And the doc's "7-13× compression" headline is asymptotic — the crate's own named falsifier compression_at_shallow_depth_is_far_below_the_headline (:744) pins ~2.1× at d=4,N=2.

TECH_DEBT — TD-SIGKER-CLIPPY-RED-ON-BASE-1. sigker fails clippy --all-targets -D warnings at signature.rs:133, and did so before this change (verified by stashing the tree and re-running against base). Not fixed here — that would widen a wiring PR — but the 3-line patch is in the entry. Root cause worth noting: sigker is workspace-excluded, so root-level cargo clippy never reaches it. The other excluded crates likely share the blind spot.

Test plan

• cargo test --manifest-path crates/sigker/Cargo.toml — 62 passed, 0 failed
• cargo fmt --manifest-path crates/sigker/Cargo.toml -- --check — clean
• cargo clippy --manifest-path crates/sigker/Cargo.toml --all-targets -- -D warnings — clean on every site this diff touches; only the pre-existing signature.rs:133 remains, tracked above
• Board files verified append-only (insertions only, zero deletions)

Not changed

examples/randomized_signature_demo.rs and cubature_vs_randomized.rs only call encode / new with no internal-field access, so they need no edit.

🤖 Generated with Claude Code

https://claude.ai/code/session_016WkNBjHc2e3zuyz9i8qJEv

---

Generated by Claude Code

[coderabbitai]

Warning

Review limit reached

Next included review available in 5 minutes.

Check out review usage here.

View limit details

Limit details: You’ve used the included review currently available. Your 87 included PR review attempts over the past 7 days set your current allowance at 1 review per hour.

Your organization has reached its usage spending cap. Adjust your spending cap in the billing tab.

Learn how review limits work.

Review configuration:

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Essentials

Run ID: d811dad2-6e17-47ae-8c91-52fb07445628

📥 Commits

Reviewing files that changed from the base of the PR and between ac9148f and 47c9c3d.

📒 Files selected for processing (4)

• .claude/board/EPIPHANIES.md
• .claude/board/TECH_DEBT.md
• .github/workflows/rust-test.yml
• crates/sigker/src/randomized.rs

---

_{Comment @coderabbitai help to get the list of available commands.}

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: serverGenReqId_377fbce7-e47f-4562-8c47-be08d0de482d)