solidcraft-upstream/docs/perf.md

Design loading

STEP assemblies: a repeated part is meshed once (2026-10-14)

The reader already converts each B-rep item once and keeps it for the instances that list it again. Every instance still became a body of its own geometry (it has to: features cut and join it where it is), and each of those bodies made its own display mesh: tessellation, seam and tangent flags, the most expensive step of opening a large assembly. A board with 60 of each resistor meshed the same resistor 60 times.

A later instance of an item now shows its first instance's display mesh, placed by the difference of the two placements: the same triangles, faces and edge flags, moved. The mesh is made once, by whichever instance is drawn first. Only a rotation and translation is shared this way (anything else meshes itself), and anything that changes what a body shows (split lines; any operation, which makes a new body) drops the link. Measurements, booleans and export read the instance's own geometry as before, so solidcraft-cli eval of an assembly is unchanged.

Two generated assemblies (SolidCraft designs with placed copies of components, exported as STEP AP214 assemblies), solidcraft-cli bench-open, release build, shared 16-core Linux box:

file	bodies	mesh phase before	after	CPU before	after	peak RSS before	after
4 parts, 500 instances (bolts with a torus ring, brackets)	901	2,215 ms	139 ms	28.4 s	4.1 s	918 MB	749 MB
10 filleted parts with 8 pins, 600 instances	5,401	1,028 ms	85 ms	16.7 s	6.0 s	1,095 MB	783 MB

Both import to the same triangle counts (4,404,412 and 4,316,412). Each instance's display plan (the GPU viewport draws bodies from plans) is still made per instance; sharing those the same way is the next step.

Designs built by agents: work per command that grew with the design (2026-10-14)

A design an agent builds over MCP is made of many small commands: a cylinder per spoke, a box per chain plate, one part built at the origin and its component moved into place. Each command did work over the whole design, so the 1,000th command cost far more than the first and a 1,300-body design took minutes to build and to show. solidcraft-cli bench-agent builds such a design (a generated bicycle: about 1,260 commands, 1,280 bodies, 1,250 features, wheels built at the origin and moved), timing each command and then what a window needs to draw its result (the world state and its display meshes), and finally opens the saved design like bench-open. The script is generated, so it needs no fixture; --scale changes the counts and --script-out writes the commands for replaying elsewhere (engine.execute one by one over the control channel reproduces the desktop case).

What grew with the design, per command:

• Moved components were placed again after every change. world_state transformed every body of a component whose occurrence is moved and gave it a new, empty display mesh, so every command meshed all those bodies again (the two wheels: 150 bodies and two tori, 130-190 ms per command). It now keeps the bodies it placed and reuses each one whose definition (same body, same display cell) and placement are unchanged.
• A free dN name searched the design once per number (next_d_name), quadratic in the number of named inputs, for every new feature input.
• Every expression evaluated every parameter (Document::eval), several times per feature command (each coordinate, each size). Parameters are now evaluated only when the expression names one.
• The parameter table's used_by searched all parameters for each one (quadratic); it is one pass now. The desktop also stops gathering expression completions (eleven parameter tables) on every change while no text box has focus.

solidcraft-cli bench-agent, release build, 16-core Linux shared with other builds (load average 15-20), one run each:

per command, by design size	before: command	before: display	after: command	after: display
200 commands (227 bodies)	6.0 ms	129 ms	1.8 ms	2.7 ms
600 commands (621 bodies)	19.9 ms	169 ms	4.9 ms	2.4 ms
1,000 commands (1,021 bodies)	45.1 ms	187 ms	9.8 ms	4.5 ms
1,200 commands (1,221 bodies)	56.8 ms	194 ms	12.5 ms	5.6 ms
whole build (1,260 commands)	241 s		16.3 s

The command column still grows (the timeline's fingerprints and parameter values are walked once per command), but slowly. solidcraft-cli eval output for the design corpus, the examples, the regression scripts, the generated bicycle and a 3,300-feature agent-built design is unchanged apart from the timings.

STEP import: shells and item solids without copies (2026-10-13)

A file that lists one product definition many times (a bolt used 70 times, a panel in several assemblies) is imported item by item, and every item's B-rep was deep copied twice before it was handed on: once by the caller of Solid::try_new, which clones the shells so that the failure path still has them, and once by the reader's own cache, which kept the solid for a later item that might list the same definition.

read_item now runs those four checks (non-empty, connected, closed, no singular vertices) on borrowed shells, so a shell that passes is moved into the solid instead of cloned. The reader keeps each prepared solid in an Arc, and an item counts as once when one representation lists it and exactly one product-definition walk reaches it: that item's solid is moved out of the cache with Arc::try_unwrap instead of copied, because no later item can ask for it. The walk follows MAPPED_ITEM targets too, so an item reached through a mapping in another product's node is still converted once - a second conversion would repeat that item's warnings. A file whose walk exceeds the work budget (MAX_CLOSURE_WORK, 65 536 nodes) falls back wholesale to the copying path, as does any file with component occurrences, which walks a definition again per instance.

Nothing observable changed. Fifteen fixtures (primitives, a void, a fillet, a revolve, colours, centimetre units, 70 duplicates of one definition, a broken file with mappings, a two-instance assembly and an expanded one; 99 bodies) give the same import dump - name, path, closed flag, file faces, face origin, topology and merged counts, bounding box, volume, area, centroid, paint, warnings and tree - apart from the per-feature milliseconds, and all fifteen STEP in/out round trips are byte-identical.

Two generated flat STEP files, timed like the table above (same binary pair, no other work running); four alternating runs each, two for the 31 MB file:

File	Baseline peak RSS	Patched peak RSS	Import
2500 boxes, 9.5 MB	150.1 MB	145.9 MB	1.81 s → 1.74 s
2500 filleted boxes, 31 MB	428 MB	433 MB	70.3 s → 71.0 s

The first row's ranges do not overlap (150.0-150.6 MB against 145.0-146.0 MB), and the saving is the shell copy that is no longer made for every item. The second file's runs land between 422 MB and 444 MB either way, so two runs each do not resolve it: there the conversion of the geometry dominates the per-item copy. An item that keeps every conversion - an assembly, or a file over the work budget - is unchanged.

Display tessellation plans (#303, 2026-10-12)

Body::display_mesh meshes from a TessPlan (crates/kernel/src/tess): every edge is sampled once and both faces use those nodes exactly (watertight by construction); planar faces are a constrained Delaunay triangulation of their edge nodes alone; faces bounded by their parameter box are structured grids zipped to the edge nodes; other four-cornered faces are Coons grids; only genuinely trimmed curved faces (holes, more than four corners) use spade's constrained Delaunay with sag and Ruppert refinement. A body the planner cannot handle is meshed by truck-meshalgo as before, which also still meshes for modelling (booleans, measure, provenance). Evaluating a plan (surface points and normals) runs on the CPU (TessPlan::evaluate) or in a compute shader (solidcraft-gpu).

Release build, Apple M4 Max, medians of five. 400-hole plate (200 × 200 × 5 mm, 806 faces), cargo test --release -p solidcraft-gpu bench_plate -- --ignored --nocapture:

Tolerance	Reference mesher	Plan	CPU evaluation	GPU evaluation	Triangles (reference / plan)
0.3 mm	23.3 ms	11.9 ms	1.1 ms	1.1 ms	20,812 / 20,812
0.03 mm	54.2 ms	31.6 ms	2.5 ms	1.9 ms	76,812 / 52,812
0.006 mm	118.9 ms	65.1 ms	5.3 ms	3.1 ms	302,412 / 104,012

Planning (edge sampling, parameter projection, triangulation) dominates; evaluation is a few milliseconds either way, so GPU evaluation saves little until planning is cheaper. The GPU timing includes upload and read-back.

Parallel STEP body preparation (2026-10-12)

Profiling an original 48-cylinder STEP fixture before changing the importer measured 3.17 ms parsing, 6.06 ms reconstructing B-reps and 45.58 ms preparing/orienting bodies. The five complete imports had a median of 59.21 ms. The expensive orientation mesh, rather than the text parser, was the first parallelization target for #287.

Desktop import now prepares consecutive independent solids in a representation with up to eight workers. Each batch contains at most 64 solids. Parsing, B-rep reconstruction, assembly traversal, first-use unit resolution, naming and result collection remain ordered. Mapped items end a batch. Fewer than four solids use the serial path; wasm always uses it. Thread creation failure falls back to serial work, and kernel/worker failures remain per-body errors reported by the importer.

Windows x64, AMD Ryzen AI Max+ 395 (16 cores / 32 logical processors), optimized ci profile, no concurrent build: three fresh test processes per fixture, five alternating serial/parallel pairs in each process. These are complete uncached kernel imports, including parsing and body preparation; fixture generation/export and later desktop display meshing are outside the timer. Both paths use the same input text and code.

Original generated fixture	Bodies	Serial median	8 workers median	Reduction
Cylinders, radius 5 mm, height 12 mm	48	57.54 ms	17.38 ms	69.8%
Cylinders, radius 5 mm, height 12 mm	512	896.31 ms	313.06 ms	65.1%
Tori, major radius 20 mm, minor radius 4 mm	48	269.05 ms	51.90 ms	80.7%
Boxes, 10 × 15 × 12 mm	512	247.97 ms	90.15 ms	63.6%
Boxes, 10 × 15 × 12 mm	4	1.49 ms	1.02 ms	31.5%
One cylinder (serial fallback)	1	1.22 ms	1.29 ms	serial fallback

There are 15 samples per cell. The single-cylinder modes run identical serial code; their 0.07 ms median difference is within the overlapping 1.11–1.57 / 1.13–1.52 ms ranges. The 512-cylinder runs ranged from 567.40–983.61 ms serial and 197.72–345.51 ms parallel; scheduling/thermal variation is material. These synthetic results do not establish a speedup for every assembly or complete desktop Open. In particular, parts containing only one solid are not parallelized across assembly nodes. Cross-part scheduling, parallel B-rep reconstruction and later display meshes remain open.

Reproduce after compiling, without other builds running:

cargo test -p solidcraft-kernel --profile ci benchmark_independent_step_solids --no-run
SOLIDCRAFT_STEP_BENCH_COUNT=512 SOLIDCRAFT_STEP_BENCH_KIND=cylinder \
  cargo test -p solidcraft-kernel --profile ci benchmark_independent_step_solids \
  -- --ignored --nocapture --test-threads=1

Use torus or box for the other fixtures and repeat in three fresh processes. SOLIDCRAFT_STEP_BENCH_FIXTURE optionally writes the original STEP for native inspection. No timing threshold runs in CI. Regressions compare serial and threaded imports with different worker counts: exact display meshes, measures, body/face source identities, names, colours/opacity, units, rotated assembly placements, duplicates and partial failures across the batch boundary. The public analytic/FreeCAD reference suite also checks STEP export/import; its current results are reported in docs/oracle.md.

Design storage: container and binary payloads (#303, phase one)

Design containers avoid formatting/parsing imported mesh coordinates as JSON numbers. The metadata serializer replaces embedded fields before creating JSON values, without cloning the full document or duplicating its schema. Binary coordinates retain f64 precision, indices retain u32, and SHA-256 deduplicates identical payloads. The reader checks archive budgets, CRCs, hashes, exact lengths, finite coordinates and index bounds.

Reproduce the storage measurements with an optimized ci IO test executable:

cargo test -p solidcraft-io --profile ci benchmark_ -- --ignored --nocapture --test-threads=1

The benchmarks generate an original 100,000-vertex/100,000-triangle mesh and a STEP containing 512 disjoint 10 mm boxes. Fixtures are written under CARGO_TARGET_DIR (or the crate's target directory). Each compares nine encoding/decoding samples; atomic-save samples include encoding, temporary-file flush, replacement and .bak creation and alternate JSON/container ordering. Documents are compared after decoding. These are storage measurements, separate from timeline evaluation or first-frame latency.

On Windows x64, based on upstream e6e0a7e before STEP parallelism (#312), three consecutive optimized runs without concurrent compilation gave the following medians of the three nine-sample medians (milliseconds, fast mode):

Fixture / operation	JSON	Container
Mesh size, bytes	15,139,115	3,600,828
Mesh encode	23.336	6.094
Mesh decode	57.175	2.962
Mesh atomic save	52.924	14.856
STEP size, bytes	1,998,793	1,950,056
STEP encode	1.507	1.473
STEP decode	1.479	1.318
STEP atomic save	7.213	7.087

Three alternating fresh-process solidcraft-cli bench-open runs of each STEP encoding returned the same 512 bodies and 6,144 triangles. Median read-plus-parse was 2.274 ms for JSON and 2.151 ms for the container; median evaluation was 218.559/236.915 ms and display meshing 203.160/197.790 ms respectively. Geometry dominates and these variable results do not establish an end-to-end STEP opening improvement. Fresh processes do not imply a cold operating-system file cache, and these are not desktop first-frame times.

Fast mode stores STEP and mesh payloads directly. Explicit "compression":"compact" compresses them at Deflate level one; it is a space/CPU tradeoff. In the initial 512-box STEP experiment, compression reduced 1,998,793 JSON bytes to 578,659 package bytes but increased median encoding from 1.68 to 5.02 ms and decoding from 1.76 to 4.18 ms. That regression is why compression is not the default. PNG/JPEG are stored directly in either policy; small parametric metadata uses Deflate level one.

Plain JSON remains a supported Save As/export format; .json saves keep JSON and scripts may explicitly select either format. The CLI detects packages before attempting JSON parsing. Desktop/browser recovery and browser downloads also use the container. See the versioned container format for layout and bounds.

This phase does not restore evaluated B-reps, skip STEP import, or make cold geometry rebuilds instant. A matching cache key is not enough to trust unvalidated geometry. The native B-rep increment described below supplies provisional display, not editable session restoration. It is separate from #287's import parallelism.

Opt-in mesh preview measurements (#303, phase one)

The phase-one desktop implementation can show an experimental saved picture while the file worker rebuilds geometry. It requires an explicit preview: true engine Save/Save As call. Ordinary saves omit display meshes to avoid the measured overhead; these measurements predate native B-rep caching. The cache contains only already warm evaluated display meshes, capped at 8 MiB, 1,024 bodies and 100,000 triangles. Larger or cold models save without it. Cached geometry is never used for editing, measurements or picking. Source/document fingerprints and payload validation reject stale or malformed caches; cancellation and document/revision checks prevent a background result from replacing newer work.

The cached picture is prepared on the file worker from the same file snapshot as the normal Open command. It adds work before reconstruction, rather than bypassing it, so measure preview readiness and full editing readiness separately. Cached display also adds file size and save work; the storage-only results above exclude that optional data.

For an engine-level save benchmark, first generate the original 512-box STEP fixture with the ignored IO benchmark, then run the warmed-session benchmark:

cargo test --workspace --profile ci benchmark_step_storage -- --ignored --nocapture --test-threads=1
cargo test --workspace --profile ci benchmark_warm_design_save_with_preview -- --ignored --nocapture --test-threads=1

Set CARGO_TARGET_DIR to the workspace's absolute target directory in both commands. The engine benchmark alternates nine JSON/container atomic saves, includes the optional display cache, and verifies both preview decoding and authoritative document identity. Three runs on Windows, with no concurrent build or test suite, measured median atomic JSON saves of 7.679/7.610/7.630 ms and container-with-preview saves of 10.319/10.633/10.541 ms. Preview decoding took 0.964/0.961/1.130 ms; the container was 2,510,490 bytes. This optional preview costs save time on this STEP fixture. An actual desktop sensor-enclosure opening on e8a8e0f prepared its saved picture in 24.320 ms; geometry reconstruction, display warmup and thumbnail preparation took 8,913.212 ms. The same preview experiment is now explicitly opt-in; full geometry readiness still requires reconstruction. For actual desktop opening phases, launch with SOLIDCRAFT_PROFILE_OPEN=1: ready_ms reports saved-picture preparation, while the normal phase log reports reconstruction, display warmup and thumbnail completion. Neither log measures the exact GPU first frame.

Native cache opening (#303)

The portable native B-rep cache is opt-in with cache: true on warm container saves, without duplicating display meshes. Ordinary saves and browser downloads default to cache: false: the measurements below do not justify imposing the additional save time and file size on every design. Desktop Open reads metadata and optional geometry through a seekable file, produces a bounded provisional picture, then resolves original sources and reconstructs the complete session before editing. This improves the opportunity for early display, not the cost of authoritative feature evaluation. An absent, invalid or incompatible cache uses the original reconstruction. Browser OPFS/downloads share the format; browser opening still rebuilds synchronously.

The reproducible engine benchmarks separate atomic saving, provisional mesh readiness, reconstruction and final display-mesh readiness:

cargo test -p solidcraft-io --profile ci benchmark_step_storage -- --ignored --nocapture --test-threads=1
cargo test -p solidcraft-engine --profile ci benchmark_native_cache_save -- --ignored --nocapture --test-threads=1
cargo test -p solidcraft-engine --profile ci benchmark_native_cache_open -- --ignored --nocapture --test-threads=1

Use an absolute CARGO_TARGET_DIR. The save benchmark defaults to the generated 512-box STEP document and accepts SOLIDCRAFT_BENCH_FILE for another public design. It alternates nine saves of each encoding, including flush/replacement/backup, and verifies document identity and native-cache presence. It never times initial fixture evaluation as save work. Open runs once per process; SOLIDCRAFT_BENCH_ENCODING selects 0=JSON, 1=container without cache, 2=native cache, 3=compact native cache. Changing only cache/key.json's application version provides a cache-miss control. Fresh processes do not imply a cold operating-system file cache. The engine callback time is provisional mesh readiness, not a rendered desktop frame; use the desktop profile log and actual screenshots to check the saved picture and completed model.

Measured 2026-10-13 on Windows 11, Ryzen AI MAX+ 395 (16 cores/32 threads), ci profile, with the neutral codec revision 75381fa, envelope SCBREP03, producer solidcraft-truck-adapter/2, and opening integration revision 3243422 on upstream 2880525 (0.6.0). These replace the earlier prototype measurements. They compare encodings on the same code and do not attribute upstream modelling improvements to the cache. Runs started after local builds and the complete CI suite finished; ordinary desktop activity remained. Save values are medians of nine alternating atomic saves. Open values are medians of five fresh processes per encoding, alternating forward/reverse order. Memory is each process's peak working set from Windows GetProcessMemoryInfo, sampled every 10 ms, then the median across those five processes; it excludes a desktop window/GPU.

The original STEP fixture is a 32 by 16 array of 10 mm cubes, 12 mm apart, exported by SolidCraft: 512 bodies and 6,144 final triangles. Its JSON embeds about 2 MB of STEP. Every encoding reopens without feature errors and with those same counts.

STEP encoding	Bytes	Atomic save, ms
JSON	1,998,826	12.616
Container, no cache	1,950,084	10.516
Native cache, fast	4,694,035	47.755
Native cache, compact	837,067	44.015

STEP encoding	Provisional mesh, ms	Reconstructed, ms	Final meshes, ms	Peak working set, MiB
JSON	none	175.842	748.911	40.102
Container, no cache	none	175.885	722.306	40.504
Native cache, fast	62.099	228.794	747.976	40.883
Native cache, compact	65.981	235.403	762.120	40.918
Native cache miss	none	169.039	647.617	40.453

The public examples/acceptance/sensor_enclosure.json script, saved as a JSON design, has 21 features, two bodies and 4,364 final triangles on this revision.

Sensor enclosure encoding	Bytes	Atomic save, ms
JSON	52,281	5.243
Container, no cache	5,545	4.969
Native cache, fast	379,095	12.646
Native cache, compact	72,603	14.111

Sensor enclosure encoding	Provisional mesh, ms	Reconstructed, ms	Final meshes, ms	Peak working set, MiB
JSON	none	3,979.063	3,984.593	70.438
Container, no cache	none	4,038.127	4,042.006	71.789
Native cache, fast	241.541	4,269.124	4,274.272	70.465
Native cache, compact	236.739	4,271.254	4,275.159	70.148
Native cache miss	none	4,136.517	4,140.903	69.941

Native display is available earlier, at the cost of encoding, file size and work before reconstruction. It does not make the full design editable earlier, and these results do not establish a memory reduction. Final-mesh readiness varied from 611 to 856 ms for STEP and 3.84 to 4.60 s for the enclosure; small differences in full-opening medians should not be read as modelling speedups. Compact storage reduces this STEP example's size and atomic-save time, but costs slightly more save time on the enclosure. The default cache: false retains the phase-one save path; compression: "compact" is an explicit space/CPU choice. Larger or unsupported geometry can exceed the documented budgets and reopen without a provisional image.

A separate desktop check used a fresh 0.6.0 application process for each enclosure encoding, with no local compilation or test suite running. The native cache's provisional picture was prepared at 429 ms and captured visibly during opening; the completed model was adopted in 6.48 s. JSON had no provisional picture and completed in 6.08 s. These are single visual checks, not repeated timing medians; the worker reported 6.10/5.66 s respectively, and external observations include UI scheduling and screenshot requests. During native opening, 94 control-channel inspections had a median round trip of 20.8 ms and a maximum of 106.5 ms (JSON: 95 inspections, 20.3/36.9 ms). The provisional picture is explicitly labelled as being checked before editing; the final frame shows the rebuilt Case and Lid and their editable feature timeline. This demonstrates early feedback, not faster completion of feature evaluation or a worst-case UI-latency guarantee.

Analytic extrusion source preparation

New-body extrusion certificates now prepare the sketch's bounded analytic segments once per certification pass instead of allocating segments and resolving their points for each generated cap boundary and side face. Saved designs run this same code during cold rebuilds. Exact containment tolerances, ambiguous-source rejection and the certification work budget are unchanged. Prepared segments borrow an immutable evaluated sketch and are not retained across edits.

The reproducible microbenchmark is:

cargo test -p solidcraft-doc --profile ci benchmark_prepared_extrude_sources -- --ignored --nocapture

Three Windows x64 optimized ci runs, during other compilation, compared the previous segment reconstruction loop and the prepared matcher in the same process. For 1,000 analytic lines and 1,000 generated face segments, median time fell from 45.90 ms to 12.30 ms (73% less, about 3.7x faster). Preparation is included and every returned source identity is compared. This measures certification only, not complete extrusion, file-open latency or Boolean operations; it does not establish an end-to-end loading speedup. Remaining cold geometry rebuild and import costs are tracked in #255; certification work is #254.

Cold geometry rebuilds after 0.3.1

Three alternating cold-process comparisons on the same Windows machine and optimized ci profile, with no concurrent build, compare the unchanged 0.3.1 source with the straight-line projection and exact surface-refinement reuse changes:

Saved design	Before, median evaluation	After	Reduction
Public sensor enclosure, 21 features, 52 KB	11.96 s	8.85 s	26.0%
Local Pi case assembly, 25 features, 2,990 bodies	14.75 s	11.57 s	21.6%

These are geometry rebuild times, not complete desktop opening times. The public fixture can be generated with solidcraft-cli run examples/acceptance/sensor_enclosure.json --save sensor.solidcraft --quiet, then measured with solidcraft-cli bench-open sensor.solidcraft. Each benchmark starts a new process; this does not measure a warm timeline-cache hit. The local assembly is private and is not part of the repository.

Straight segments now use guarded algebraic closest-point projection instead of sampling 65 points and running Newton iterations. Other curve types retain their existing behavior. Curved-face refinement reuses derivative and midpoint evaluations with exact finite input keys and bounded, local caches. It keeps the original tolerance, seam and singularity checks, refinement passes, and triangle limits.

Identity assembly placements retain the definition body's shared display-mesh cache. Changed placements continue to build separate meshes in their own coordinates. Face-name preparation remains eager during placement: deferring it merely shifted expensive work to thumbnails and the first face query, so that approach was excluded.

Both saved fixtures retain exactly the same body names, face/edge/vertex/shell counts, bounds, volumes, and surface areas. Focused tests compare cached and uncached refinement coordinates and normals bit for bit on periodic, singular and transformed surfaces.

For desktop phase profiling, set SOLIDCRAFT_PROFILE_OPEN=1 before launching the app. Interactive Open reports geometry, placement, tessellation, thumbnail and total worker times to stderr without logging the source path. Thumbnails were about 0.1 seconds of the slow assembly's open and were not the main bottleneck. Further cold Boolean and import costs remain; these changes do not make all documents open instantly.

Prepared assembly geometry and selection

Prepared world geometry is now keyed by the document, evaluated model and exact placement context instead of the session revision. Selecting a face or edge, changing visibility, or moving the view no longer discards already prepared assembly bodies and meshes. Geometry edits, pending occurrence moves, authoring context changes, undo/redo and timeline rollback still invalidate the cache. Scratch sessions share prepared geometry while their inputs match, and adopting a computed session retains its warm cache.

On the private 2,990-body Pi assembly, a matched Linux desktop control probe reduced an edge-selection round trip from 9.24 s to 0.57 s. Initial opening stayed essentially unchanged (33.54 s to 33.32 s); this fixes a separate selection stall, not cold geometry evaluation. Three fresh Windows desktop runs with the corrected cache measured a median 15.87 s to ready: 13.04 s geometry, 2.20 s placement, 0.53 s meshing and 0.07 s thumbnails. These used the optimized ci profile and actual Vulkan rendering, with isolated settings and caches. Linux and Windows timings are separate measurements and must not be compared as a speedup.

solidcraft-cli bench-open design.solidcraft reports JSON lines for file reading, JSON decoding and migrations (including input-name repair), timeline evaluation, individual features, and world placement/display meshing. --parse-only stops before geometry evaluation. The parse timer includes document validation and repair, not just the JSON parser.

Windows x64, optimized ci profile, measured during other builds:

Design	Parse/repair	Evaluate	Display meshes
19-feature enclosure, 34 KB	1.2 ms	14.9 s	6 ms
25-feature assembly, 15.2 MB, 2,990 bodies	21 ms	12.5 s	2.1 s
Generated 1,000 independent boxes, 185 KB	5 ms	0.82 s	0.15 s
Generated 5,000 boxes in five grids, 10 features	0.45 ms	17.6 s	0.81 s

The enclosures spend most of evaluation in joins and cuts, not parsing. Designs save feature definitions; opening them rebuilds their geometry. Those times remain after the repair fix.

A separate generated timeline with missing input names reproduced minutes in the reader:

Timeline (geometry rolled back)	Parse/repair before	After
2,000 boxes, 6,000 unnamed inputs	78.1 s	10–20 ms
5,000 boxes, 15,000 unnamed inputs	interrupted after 131 s, still repairing	26–46 ms

Previously, each missing input restarted the d1, d2, … search, scanning the document for every candidate. This approaches cubic work as unnamed inputs accumulate. Loading now reserves names once and assigns missing names with an advancing cursor. Existing names, names on later features, parameter collisions and reuse of surplus names retain the previous behavior. Regression tests compare the old and bulk repairs and load a 5,000-feature unnamed timeline.

Viewport performance

Measured with the control channel (ui.inspect frame_ms: CPU time of one UI frame, which includes picking under the cursor) on the sample plate patterned 20 × 20 (400 bodies, about 600k triangles), 1600 × 1000 window under Xvfb, release build. Hover: 60 cursor moves across the model; orbit: a right-drag.

	hover median	hover p90	orbit median
before	18.7 ms	37.5 ms	21.0 ms
after	1.9 ms	2.4 ms	2.6 ms
after, Browser hidden	0.4 ms	—	0.4 ms

What changed:

• Picking culls by bounding box. Per-body boxes are cached until the model changes. The face raycast skips bodies whose box the ray misses or enters behind the nearest hit so far. Edge and vertex tests skip bodies whose box, projected on screen, is not under the cursor.
• Picks are reused while the cursor, camera and model are unchanged, and no picking runs while the view moves (orbit, pan, zoom).
• The model extent is cached (it sets the view's depth range and was recomputed from every mesh several times per frame).

With 49 bodies the frame was already 2.8 ms (hover) before these changes. The remaining frame time with the Browser shown is mostly the Browser listing every body row. GPU ID-buffer picking was not needed at this size.

Browser

Same model and method (400 bodies, Bodies folder open, 60 hover moves over the viewport):

	median	p90
Browser shown, before	1.95 ms	2.71 ms
Browser shown, after	0.72 ms	1.19 ms
Browser hidden	0.30 ms	0.72 ms

What changed: rows scrolled out of view only reserve their height. They register no widgets and paint nothing (no text layout, icons or eye/fold buttons), so the cost follows the rows on screen rather than the rows in the tree. Scrolling, Find in Browser and drag and drop still see every row's place.

Recompute performance

solidcraft-cli bench-edit <script> <param> "(<old>) * 1.02" builds a design from its script (timing the build), then times a parameter edit, setting the parameter back to its old expression, undo and redo. Each cell is before → after the evaluation cache, with the number of features recomputed in parentheses. Release build, on a machine shared with other build jobs (load average about 45; the times vary by ±50 % run to run, so compare the recompute counts and the zeros). "Early" edits a parameter of the first sketch or feature, "late" one of the last features.

Design	Edit	Build	Edit	Set back	Undo	Redo
sample plate, 20 × 20 pattern	early (width)	7587 ms (10) → 1764 ms (10)	7969 ms (10) → 1691 ms (10)	7557 ms (10) → 0 ms (0)	6534 ms (0) → 0 ms (0)	4570 ms (0) → 0 ms (0)
sample plate, 20 × 20 pattern	late (d9)	7870 ms (10) → 1699 ms (10)	3944 ms (1) → 812 ms (1)	4039 ms (1) → 0 ms (0)	1464 ms (0) → 0 ms (0)	1398 ms (0) → 0 ms (0)
oracle 58 phone stand	early (d1)	163 ms (6) → 140 ms (6)	228 ms (5) → 140 ms (5)	171 ms (5) → 0 ms (0)	158 ms (0) → 0 ms (0)	130 ms (0) → 0 ms (0)
oracle 58 phone stand	late (d4)	144 ms (6) → 195 ms (6)	118 ms (1) → 118 ms (1)	145 ms (1) → 0 ms (0)	149 ms (0) → 0 ms (0)	187 ms (0) → 0 ms (0)
oracle 59 enclosure base	late (d8)	11582 ms (11) → 11179 ms (11)	5772 ms (1) → 5742 ms (1)	5606 ms (1) → 0 ms (0)	6023 ms (0) → 0 ms (0)	6119 ms (0) → 0 ms (0)
oracle 60 pipe flange	early (d1)	8137 ms (7) → 8348 ms (7)	7569 ms (6) → 10143 ms (6)	8354 ms (6) → 0 ms (0)	8607 ms (0) → 0 ms (0)	10034 ms (0) → 0 ms (0)
oracle 60 pipe flange	late (d6)	7661 ms (7) → 11990 ms (7)	67 ms (1) → 361 ms (1)	69 ms (1) → 0 ms (0)	84 ms (0) → 0 ms (0)	66 ms (0) → 0 ms (0)
oracle 61 shaft keyway	early (d1)	1723 ms (6) → 3016 ms (6)	1785 ms (5) → 2745 ms (5)	1760 ms (5) → 2323 ms (5)	1667 ms (0) → 0 ms (0)	1719 ms (0) → 0 ms (0)
oracle 61 shaft keyway	late (d4)	1905 ms (6) → 2247 ms (6)	1734 ms (1) → 2101 ms (1)	1682 ms (1) → 0 ms (0)	1719 ms (0) → 0 ms (0)	1876 ms (0) → 0 ms (0)
oracle 62 heat sink	early (d1)	382 ms (6) → 643 ms (6)	362 ms (5) → 697 ms (5)	233 ms (5) → 568 ms (5)	245 ms (0) → 0 ms (0)	360 ms (0) → 0 ms (0)
oracle 62 heat sink	late (d5)	479 ms (6) → 688 ms (6)	270 ms (1) → 657 ms (1)	321 ms (1) → 0 ms (0)	413 ms (0) → 0 ms (0)	283 ms (0) → 0 ms (0)
oracle 63 hinge knuckle	early (d1)	1305 ms (8) → 1671 ms (8)	1329 ms (7) → 2397 ms (7)	1971 ms (7) → 0 ms (0)	1705 ms (0) → 0 ms (0)	1321 ms (0) → 0 ms (0)
oracle 63 hinge knuckle	late (d4)	1430 ms (8) → 2072 ms (8)	1387 ms (1) → 2073 ms (1)	1480 ms (1) → 0 ms (0)	1192 ms (0) → 0 ms (0)	1035 ms (0) → 0 ms (0)

What changed: every feature evaluation is cached by a key made of the feature's own inputs (definition and parameter values), the document context it may read (construction planes, sheet metal and plastic rules, components) and a digest of the model it starts from, in which each body is identified by the evaluation that produced it (bodies a feature passes through unchanged keep their identity). The same feature on the same inputs is never evaluated twice: setting a value back, undo, redo, a preview followed by the same command, and the repeated evaluations while a script builds the design all reuse earlier results. The cache is shared by a model and its copies (previews, scratch sessions), and keeps the last 4096 results within 64 MiB.

An edit still recomputes the edited feature and every later feature whose starting model changed; in these parts every later feature works on the edited body, so early edits recompute almost everything, as they must. cached_evaluation_matches_from_scratch checks that cached and from-scratch evaluation give the same models (bodies, volumes and areas to 1e-6, topology, errors) through edits, suppression, roll-back, undo and redo. Setting a sketch dimension back can still recompute (oracle 61, 62): solving the sketch moves its points by rounding noise, so its inputs are not exactly the old ones. Oracle 59's first parameter edit fails (a recipe expression the bench's * 1.02 can't scale) and is left out.

Prism pairs, body measurements and the evaluation cache budget (2026-10-13)

A design whose features repeat one shape (a hole pattern, a plate cut by many tools, a part built by one growing body) pays several per-feature costs that a rebuild was paying again and again.

A body's size is measured once, not once per call. Every kernel tolerance is a fraction of the body's bounding-box diagonal, and Body::size() walked every vertex of the solid to work it out. Body::new already reads those vertices to orient the body, so it keeps the diagonal it measured there; a body built by hand, or a mesh body, still measures once.

The same volume is not measured twice. A boolean measures both operands' volumes to bound its result, and the body it returns becomes the next feature's operand, so a timeline of n cuts meshed every intermediate body twice. Body::volume_at(tol) keeps the last measurement for the solid, shared by its clones (they share the solid), and the bounds ask at the same tolerance every time.

A prism pair is decided in 2D before either operand is meshed. The general boolean builds a refined mesh of both operands before it can decide anything, and that mesh is most of its cost: a pattern, or a sketch with several profiles, rebuilds the whole target's mesh once per tool. A plate and a tool through it are prisms along one direction, which prism::prism_boolean decides exactly from their 2D sections - and that route ran after the meshes. A cut or an intersection now asks it first and returns without building them. A pair the route cannot answer takes the old path unchanged, meshes and all; a union still asks after the sampled shortcut. One answer is sent back to that shortcut on purpose: when the 2D route returns one of its operands (a tool that misses its target, or covers it whole), the sampled path returns that very body, so an untouched body keeps its identity in the timeline (ops::prism_fast_path_tests).

A placement reads its orientation from its matrix. Placing a body (a pattern instance, an aligned copy) used to mesh the whole copy to decide which way its faces point and to count the faces its coarse mesh misses. A rigid placement leaves both as they were; a mirrored one (a negative determinant) turns the solid inside out. Both are now read from the matrix, so only the geometry is transformed.

The evaluation cache is bounded by the bytes it keeps. It held the last 4096 results, and a result is as big as the model it holds, so a long editing session kept every superseded state: nothing dropped the entries nothing referred to any more. It now also holds at most 64 MiB, oldest first, and always keeps the newest result, so a single evaluation is never thrown away in the same breath.

Release solidcraft-cli, eval <design> --quiet, wall time of the whole process and peak RSS from /usr/bin/time -v, alternating the baseline and patched binaries, median of 12 runs each (8 for the 400-hole plate and the vent plate), no other build or test running (Fedora Linux, AMD Ryzen 7 5700X, 8 cores / 16 threads, 32 GB):

Design	Baseline	Patched	Change	Baseline peak RSS	Patched peak RSS
400 × 400 × 6 mm box, one sketch with 100 circles cut through it	1.24 s	0.83 s	-33%	151 MB	119 MB
The same plate with 400 circles in the sketch	5.67 s	4.04 s	-29%	556 MB	481 MB
Vent plate, a 10 × 8 pattern of one cut	0.10 s	0.08 s	-20%	22 MB	22 MB
14 × 14 pattern of a filleted box (196 copies)	6.70 s	6.52 s	-2.6%	67 MB	67 MB
60 sketches, 60 extrudes into one growing body	8.80 s	8.88 s	+0.9%	225 MB	225 MB
examples/acceptance/sensor_enclosure.json	3.81 s	3.30 s	-13%	198 MB	196 MB

The hole plates are what the prism route is for: one boolean whose tool is a prism with 100 or 400 sections, in a sketch with that many circles, so the tool's mesh alone is what the route saves. A variant of the patched tree with the route asked in the old order - after the operands' meshes - timed the 100-hole plate at 1.27 s and 157 MB (range 1.22-1.30 s), the baseline's own numbers (1.25 s and 156 MB over the same runs, range 1.21-1.33 s). The two plate rows, and the peak they lose, are the prism route.

The last three designs are the other side: their booleans are not that case. The 196-copy design and the 60-sketch design are within 2.6% and 0.9% with overlapping ranges. The sensor enclosure's medians differ by 13%, but its runs land anywhere between 3.0 s and 5.0 s and the two ranges overlap (3.01-4.97 s against 3.00-4.89 s), so its row is not a result either way and is left as measured. The body measurements, the placement and the cache budget are not what these rows measure: they stop work the corpus repeats too rarely to see (a hidden vertex walk, an operand meshed twice for one number, an orientation mesh per placement, a cache with no byte bound).

No model changed. The evaluation dump of 21 designs (10 to 400 circles in a plate, 16 × 16 to 31 × 31 box grids with and without fillets, 2500-body grids, assemblies of 40 to 500 bodies, 30 and 60 sketches, a 196-instance pattern, the bundled sample and the sensor enclosure) is the same model before and after - bodies, volumes, areas, centres, bounding boxes, topology counts and warnings - apart from the per-feature milliseconds. The kernel's test suite passes, including four new tests for the prism route and its identity rule.

Reproduce after compiling a release build, without other builds running: time target/release/solidcraft-cli eval <design> --quiet with /usr/bin/time -v, alternating the baseline and patched binaries, and compare the middle of the runs. The corpus (20 generated designs: hole plates, box grids, assemblies, many-sketch parts) and the measuring scripts are in https://github.com/sahenjp/solidcraft-perf under perf/.

Desktop loading and recent thumbnails (2026-10-10)

The Arduino Uno enclosure (19 features, 34 KB) and Raspberry Pi 5 case (24 features, 61 KB) each decode in about 1 ms. A release build spends roughly 12.1 s and 8.2 s, respectively, evaluating their geometry on this Windows machine. Changing the document encoding does not eliminate these boolean operations.

Interactive Open, recent-file clicks, drag/drop and STEP/mesh insertion now execute on a worker, including preparation of the world-space bodies and display meshes. The existing design stays available until the result is ready; a failed or cancelled job cannot replace it. Insertion preserves the file, saved state, earlier undo history and view settings. Timeline copies share their lazy mesh cell, and adopting a prepared session keeps its warm world-space cache instead of transforming and tessellating its components again on the UI. A STEP body's missing-face check reuses its orientation tessellation when the tolerances match and its orientation was not reversed.

Live desktop tests (ci profile, with other builds/tests running) opened the Arduino and Pi case in 15.2 s and 10.5 s. During those jobs, median ui.inspect round trips were 20.5 ms and 8.3 ms; the slowest were 33.7 ms and 28.6 ms. This fixes the window freeze, not the remaining cost of cold geometry evaluation.

Recent thumbnails are 240 × 150 PNG files in SolidCraft/thumbnails under the platform cache directory (LOCALAPPDATA on Windows). Their keys include the source path, file size and modification time. Open and Save render from the model already in memory. Cached recents decode the PNG without reading or rebuilding a document, and run before cold recents and sample thumbnails. Missing or invalid pictures fall back to a background rebuild. No thumbnail or base64 image is added to the design JSON. A regression test verifies the source bytes remain unchanged, a cached picture works even without a parseable source file, and changing the source invalidates the cache.

solidcraft-cli bench-open also accepts STEP/IGES. For exchange files its parse phase includes constructing and checking the imported B-rep, not merely decoding STEP syntax.

The live test loaded both cached recent cards in 21 ms and inserted the 2 MB lightweight Pi STEP (27 bodies) into the case in 279 ms; undo restored the original two bodies. The full 77 MB board STEP (2689 solid definitions, 45,689 faces) still had not completed after 300 s. Its window continued handling inspection and Cancel, and cancelling preserved the case. Complex STEP construction remains an outstanding kernel performance problem.

Sketch solver and drags (2026-10-11)

Unknowns the constraints pin down one after another from the constants (each one's own rows of full, well-conditioned rank, reading nothing but constants and unknowns pinned before it) are eliminated exactly: values the constructive seed placed and whose constraints hold there stay constants while Levenberg–Marquardt solves the rest, and the rank analysis (degrees of freedom, determined geometry) runs on the remainder only. A component tied to no constant takes one point as a constant first (its translation). All the held solves of one live drag frame share a 1M decomposition work budget (about 16 ms; a solve's own, 20M, is about 0.3 s), and a drag's steps are a quarter of the smallest held size instead of an eighth, except once a seed ran out of budget: the rest of that drag goes in eighths, as before, and continuation keeps the branch the seed no longer places.

cargo run --release -p solidcraft-sketch --example solve_bench, macOS arm64, load average 45–75 (best of 5):

Grid (origin corner fixed, rows horizontal, columns vertical, every edge a distance)	Start	Before	After
10 × 10	at solution	0.36 ms	0.21 ms
10 × 10	±1 mm	0.34 ms	0.21 ms
15 × 15	at solution	1.41 ms	0.41 ms
15 × 15	±1 mm	1.39 ms	0.40 ms
20 × 20	at solution	4.06 ms	0.72 ms
20 × 20	±1 mm	4.44 ms	0.73 ms

The seed constructs every grid point, so LM gets no variable (a test counts them). Before, the time was the dense rank analysis of the whole grid. A 100-rectangle strip with nothing fixed, solved whole from a moved corner, still costs 20–27 ms: nothing is constructed there and LM iterates on every point with dense normal equations (a sparse solve would be the next step).

Drag frames, sketch.drag + world_state() per frame (60 frames of 0.5 mm, best of 5; the "before" column is one run of the same bench on the previous code):

Sketch	Points	Scripted ms/frame before → after	Live ms/frame before → after	drag_end ms before → after
100 rectangles	401	2.77 → 2.37	1.13 → 1.02	1.44 → 1.37
100 rectangles, rolled back under an extrude	401	2.49 → 2.33	1.13 → 0.97	1.42 → 1.28
100-rectangle connected strip	401	100.9 → 2.32	1.32 → 1.20	53.6 → 1.41
100-rectangle strip, rolled back under an extrude	401	203.8 → 2.23	1.31 → 1.18	54.0 → 1.42
rigid truss of 40 triangles, pinned at one end, dragged at the other	42	— → 6.8	— → 6.3	— → 0.59

The strip floats (nothing fixed), so its full solves (scripted frames, drag_end) ran the dense rank analysis on 800 variables; with one corner taken as a constant all of it is eliminated. A live frame's held solve was already cheap: the seed places the strip and LM returns at once. Of a live strip frame, the seed reads the constraints in about 0.15 ms, so a per-gesture cache of that reading is not worth its invalidation (its signs depend on the reference positions, which change every step).

Portable thumbnail save cost (#303)

Portable thumbnail generation requires explicit cache: true. Ordinary saves and browser downloads default to cache: false, so they incur none of the thumbnail work or added bytes measured below.

Measured 2026-10-13 on the same Windows machine and ci profile as above, but with SolidCraft 0.7.0 on both sides: d3d215b before the thumbnail change and thumbnail revision e180cdc after it. These measurements precede the integration of upstream 9f88537; later geometry changes are not included in these timings. A preserved executable of the baseline and the new executable ran the existing benchmark_native_cache_save on the same STEP and sensor documents. Three paired batches alternated before/after order; each batch performed nine alternating atomic saves per encoding. The values below are medians of the three batch medians, not a percentile of all individual saves. No compilation, test suite or desktop model was running during these measurements.

512-cube STEP encoding	Before PNG (ms)	After PNG (ms)	Container bytes before → after
Plain JSON	10.334	9.676	1,998,826 → 1,998,826
Container, cache: false	8.850	9.013	1,950,084 → 1,950,084
Native cache, fast	45.323	53.473	4,694,035 → 4,720,385
Native cache, compact	42.739	48.643	837,067 → 863,417

The PNG itself is 25,916 bytes; its key and ZIP records bring the increase to 26,350 bytes. Fast-cache batch medians ranged from 41.585–47.779 ms before and 53.212–56.556 ms after. Compact ranged from 36.672–43.040 ms before and 47.688–51.724 ms after. The approximately 8 ms fast-save increase is a measured cost on this example, not a general latency bound. JSON bytes and document/source blob bytes were compared across both implementations and remained identical.

The sensor enclosure deliberately receives no portable PNG: its projected triangle boxes total 4,293,457 supersampled pixels, above the four-million work budget. The native fast container remains 379,099 bytes. Its save batch medians are 14.714 ms before (12.187–14.956) and 13.728 ms after (13.439–14.648); these vary within the observed run ranges and do not establish a speedup. Home retains its existing thumbnail-generation fallback for this case. A small triangle count alone does not guarantee eligibility, because large overlapping faces cost more.

Actual desktop check: copy the STEP container to a new public temporary path, start SolidCraft 0.7.0 with that path in Recent and an empty local thumbnail cache, and show Home without opening the design. The portable picture was ready in one observed 47.6 ms check; no local PNG was created and the active document remained empty. The screenshot shows the transferred file's picture while built-in sample pictures are still loading. This is one functional visual check, not a first-frame benchmark. The separate counting-reader regression proves source-blob deferral.