[meros]

In short: every curve evaluation allocated a small list for its result, and the curve tessellation runs millions of them. The hot caller now passes a fixed buffer instead. Results do not change. On the loft boolean of #948, user instructions fall by about 5 % and cycles by 3 to 4 %.

What is wrong

BSplineCurve::evaluate returns a fresh Vec<Vec3> that nurbs_crv::crv_derivs allocates (Vec::with_capacity(der + 1)) on every call. On the #948 loft (20 sections, main d4940b63e) about 180 profile samples are that allocation. Almost all calls come from FindCurveDistanceExtrema::evaluate (SmFindCurveDistanceExtremaEFO::Evaluate): two evaluations of 3 entries per step of every distance solve in the curve-cache constants. The C++ evaluates into the caller's stack array.

Allocation	Rule case	Replaced by
Vec<Vec3> per curve evaluation in the distance solve	output buffer	two [Vec3; 3] on the stack

Change

• nurbs_crv::crv_derivs_into(cur, u, side, out) writes C^(k)(u) into the caller's slice; crv_derivs allocates and calls it.
• BSplineCurve::evaluate_into holds the old evaluate body (clamping, zero-tangent repair) over a slice; evaluate allocates and calls it. For num_derivs >= GW_MAX_DERIV it returns NLib's order error as before, without allocating an output of that size.
• Curve::evaluate_into sends the B-spline kinds (B-spline, hyperbola, ellipse, parabola) to it. Other kinds fall back to evaluate and fill missing entries with Vec3::UNDEF, which is what at() read for them.
• FindCurveDistanceExtrema::evaluate (in curve_cache.rs) uses two [Vec3; 3] buffers. This is a two-line change in curve_cache.rs. The curve-cache memo (perf: memoise curve caches in the curve value (outputs bit-identical) #992) is already in main and this branch is rebased on it; perf: keep each surface's boundary curves for CoincidenceCheck (outputs bit-identical) #1031 does not touch curve_cache.rs, and git merge-tree is clean against it.

Test plan

toebox_loft <sections> 1 full (the #948 driver, a local example, not in this PR), release build, main b206d3f5a against this branch, 3 interleaved runs each, taskset -c 0-3, perf stat -e instructions:u,cycles:u,task-clock, medians (load 1 to 3):

Sections	Metric	main	branch	Change
20	instructions (user)	54.87 G	52.08 G	-5.1 %
20	cycles (user)	10.84 G	10.48 G	-3.3 %
40	instructions (user)	167.59 G	158.34 G	-5.5 %
40	cycles (user)	33.53 G	32.21 G	-3.9 %

Profile samples (frame-pointer builds, perf record -F 1000, 20 sections, 2 runs each): malloc/free leaves under FindCurveDistanceExtrema::evaluate fall from 12 and 17 to 0 and 0. All non-timing output (counts, volumes and bounding boxes to 17 digits) is identical across all runs of main and the branch, for both inputs.

Checks

• cargo fmt --all --check: pass
• cargo clippy --workspace --all-targets -- -D warnings: pass
• cargo test -p solidcore: 2275 passed, 0 failed. New test: evaluate_into equals evaluate bit for bit, errors included, for 0 to GW_MAX_DERIV + 1 derivatives, both sides, with and without the zero-tangent repair, on two curves (one with a zero first derivative, so both repair branches run), and an empty output is an error.
• cargo check -p solidcore --target wasm32-unknown-unknown: pass
• difftest --unit, all failed=0 deferred=0 max_rel_err=0e0:

PASS nurbs.nurbs_crv cases=4113 passed=4113 failed=0 deferred=0 max_rel_err=0e0
PASS nurbs.b_spline_curve_p1 cases=2216 passed=2216 failed=0 deferred=0 max_rel_err=0e0
PASS curve.curve_p2 cases=3029 passed=3029 failed=0 deferred=0 max_rel_err=0e0
PASS curve.curve_cache cases=642 passed=642 failed=0 deferred=0 max_rel_err=0e0

Not run

• difftest --green (the CI gate) did not run locally; CI runs it.

What follows

• Based on main b206d3f5a; git merge-tree is clean against the current main 4fa6661e2.
• Stacked on this branch: perf/continuity-evaluate-into (evaluate_continuity into stack buffers; separate draft).
• A 4-wide curve scratch table for cubic curves was tried and is not proposed: instructions -5.4 %, but cycles +6 %.

Refs #948.

[bherbruck]

Approved by the automated first-pass review: CI green, mergeable, no CI/build/dependency file changes, no red-flag patterns in added code (process/network/env/unsafe outside FFI/blobs/hidden unicode), 0 known-mismatch entries added. Logic is gated by CI and the difftest merge queue.