[meros]

In short: the kernel makes many short-lived copies of spline curves and surfaces, and each copy duplicated the full control-point grid although neither copy changes it. The grid is now shared between copies and duplicated only when one side writes to it. Results do not change. On the loft boolean of #948, user CPU cycles fall by 2.4 % (20 sections) and 3.1 % (40 sections). This changes how every spline curve and surface stores its data.

What is wrong

Two copy sites on the #948 loft (20 sections, 7952 profile samples on main d4940b63e) copy whole NURB nets that neither side writes:

Site	Samples	Why it copies	Why it cannot borrow
SurfPtSolveStepper::find_boundary_solution	61	IsoCurve::from_surface(surface.clone_for_evaluation()?): the iso curve owns its surface	Curve is a 'static enum that owns its geometry; Arc<surface::Surface> would make Curve !Send (the geometry Surface holds a RefCell through CrvOnSurf)
curve_on_classify -> Brep::clone_sharing_surface_caches	98 (51 surface copies, 38 curve copies)	a private B-rep snapshot that lazy UV and gap computations write to; read-only callers discard the writes	the snapshot is written (UV trims, gap caches, owner scopes), so both copies survive and one is mutated

In both, what is copied is mostly NLib storage (srf_basic::Surface net and knots, crv_basic::Curve control points and knots), and nobody writes it.

Copy	Rule case	Replaced by
NURB net and knots in every clone	shared, rarely mutated: Arc + Arc::make_mut	shared until one side writes
compress of storage that is already compact	read-only	keeps the shared storage (same values)
knot fix-up of the copy constructor	real write	copies only the knot vector it writes

Change

• srf_basic::Surface holds net, u, v as Arc<Vec<..>>; crv_basic::Curve holds pw, u the same way. Every write goes through Arc::make_mut. The fields were already private to their modules, so all writes are in these two files.
• compress returns early when the storage is already exactly the indices in use (then it would build the same values).
• compressed_copy (the one-pass copy from perf: stop copying NURB surfaces for out-of-bounds solves and tolerance tests (outputs bit-identical) #1006) returns self.clone() when the storage is already compact, so the copy constructor shares the net too.
• Nothing else changes: Clone, PartialEq and Debug behave as before on the values. Arc of plain data keeps both types Send + Sync.

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	cycles (user)	10.84 G	10.58 G	-2.4 %
20	instructions (user)	54.87 G	53.66 G	-2.2 %
40	cycles (user)	33.53 G	32.50 G	-3.1 %
40	instructions (user)	167.59 G	163.23 G	-2.6 %

Per-site samples (frame-pointer builds, perf record -F 1000, 20 sections, 2 runs each):

Site	main	branch
iso-curve copy in find_boundary_solution	44, 29	0, 0
clone_sharing_surface_caches snapshot	53, 41	31, 38

The rest of the snapshot is the copy constructor's knot fix-up (fixup_knot_vector, upstream work) and the topology arrays. 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: 2276 passed, 0 failed. Two new tests: a clone shares the storage, compress and compressed_copy of compact storage keep it shared, and a write unshares only what it writes, leaving the original unchanged (surface and curve).
• cargo check -p solidcore --target wasm32-unknown-unknown: pass
• difftest --unit, all failed=0 deferred=0 max_rel_err=0e0:

PASS nlib.srf_basic cases=3220 passed=3220 failed=0 deferred=0 max_rel_err=0e0
PASS nlib.crv_basic cases=3417 passed=3417 failed=0 deferred=0 max_rel_err=0e0
PASS nurbs.nurbs_srf cases=3092 passed=3092 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_surface_p1 cases=2120 passed=2120 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 surface.surface_point_solve_p3 cases=1365 passed=1365 failed=0 deferred=0 max_rel_err=0e0
PASS query.face_p2 cases=1401 passed=1401 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

• This PR changes the storage of every NURB curve and surface in the kernel. difftest --green (the CI gate, every green and partial unit) did not run locally; only the 9 units above ran. CI runs the full set.
• No memory measurement: shared storage lowers peak memory, but I did not measure it.

What follows

• Based on main b206d3f5a (after perf: stop copying NURB surfaces for out-of-bounds solves and tolerance tests (outputs bit-identical) #1006, perf: memoise curve caches in the curve value (outputs bit-identical) #992 and perf: build a missing polar box without copying the surface cache in AdvSurfaceIntersector (outputs bit-identical) #1033); git merge-tree is clean against perf: keep each surface's boundary curves for CoincidenceCheck (outputs bit-identical) #1031 (dd47b13b9) and the current main 4fa6661e2.
• The copy constructor's knot fix-up could skip the knot copy when it changes nothing; not done (small).

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.