[aceRage]

Package 4B-2 of the Orca upstream catch-up (tests/orca_port_plan_remaining.md, section 3 4B-2): parallel vendor loading and startup progress. 2 upstream PRs ported: OrcaSlicer#13667 only its splash half (its PresetBundle half was already on main), OrcaSlicer#14155 adapted. 0 skipped. 2 Edge commits (test hooks and tests).

Bottom line: no change to how presets load. Parallel vendor loading, the ordered merge, the held-open splash and the CBOR cache came in with perf/startup on 2026-09-13 (17206cb, bc3b7dd, 11e73c6). That leaves this PR with the splash phase lines, a progress bar, and the tests and benchmark the plan asked for. The cache default is untouched (the owner decides that in 4B-3a).

Base: origin/main 0b3af03. Main has not moved since.

Ports

Upstream	What changes	Status
OrcaSlicer#13667 perf: speed up startup and show progress in splash screen (5a136a25d1)	PresetBundle: OrcaFilamentLibrary first, other vendors in parallel, merged in order. GUI: the splash stays up until the main window shows and names each startup phase	PresetBundle half already on main (17206cb stages A/B/C match upstream steps 1/2/3; bc3b7dd adds parallel parsing per file plus the cache; 11e73c6 holds the splash open). Splash phase lines Adapted (below)
OrcaSlicer#14155 Add startup progress bar to loading screen (8d84204aeb)	A progress bar on the splash, stepped 5/30/70/85/95	Adapted (below)

Deviations from upstream, and why

1. Phase lines go through g_active_splash, not the local scrn. Our splash can close itself on user input during a wxYield, so scrn may already be gone. There's no extra wxYield: SetText repaints with Update(), and the animation tick puts the line back every frame. Chinese keeps its double dots. "Loading current preset" reuses our already translated "Loading current presets".
2. The bar is drawn into the frame bitmap. Upstream draws it in a paint handler; ours is drawn in SetText(text, progress). AdvanceFrame() reapplies it every animation frame (progress -1 keeps the last value), and the bar never goes back.
3. Bar shape and colour. It's a slim rounded track under the loading line, inset clear of the card's rounded corners. Upstream's full-width bottom strip would be clipped by our rounded window shape. It's filled in the wordmark red instead of Orca teal.
4. No SetText(100) before Destroy(). close_startup_splash() takes the splash down right after the main window shows, so a 100% frame would never be seen.
5. New untranslated strings: "Loading printer & filament profiles", "Creating main window", "Showing main window". They show in English until the next gettext run.

Edge commits

• PresetBundle::set_sequential_load_override / set_parse_cache_override (+ sequential_load(), parse_cache_enabled()). The two function-local statics that read ORCA_PRESET_LOAD_SEQUENTIAL / ORCA_PRESET_CACHE now go through these. With no override set they read the environment exactly as before. They exist so one test process can compare the load modes.
• test_preset_load_full_set.cpp:
  • [.][Preset][PresetLoadFullSet] (determinism, hidden). It copies the whole shipped resources/profiles (59 vendor manifests) into a scratch data dir's system/ and adds user presets: one on a BBL system filament, a detached parent in base/ with a child of it (Fix user presets built on a detached preset vanishing #653), a process and a printer. Then it runs load_presets() five times: sequential with no cache (the reference), parallel with a cold cache, parallel with a warm cache, and parallel with no cache twice. It compares, line for line, every preset of all three collections in order (name, alias, inherits, vendor, system/default/visible, setting/filament/base ids, renamed_from, FNV-1a hash of the resolved config), the filament alias table (every holder of each alias, in lookup order), the vendor/model/variant list, and the selection after load_selections. It's hidden because it takes about 144 s (the 13k-file copy plus hashing); run it with "[PresetLoadFullSet]". The small synthetic [PresetLoadDeterminism] cases still guard the default run.
  • [.][Benchmark][PresetLoadBenchmark] (hidden): times load_presets() in four modes, ORCA_BENCH_RUNS runs each (default 5).

Conflicts with our own work

• Locale fix perf(presets): avoid process-wide locale locks while parsing preset JSON (from Orca #15943) #222: the vendor-parallel workers each install a thread-local C numeric locale (CNumericLocalesSetter in load_vendor_configs_from_json), and so do the per-file parse chunks. TBB nested stealing only nests setters on one thread, which the thread-local depth counter handles. Unchanged here.
• CBOR cache Add atomic file writes for config, presets, and the CBOR cache #255: unchanged. The benchmark shows the warm cache is slower than no cache on this set (see Timings), which agrees with the 4B-3 design note. The default stays as it is.
• Inherits / aliases / duplicate detection / Fix user presets built on a detached preset vanishing #653 detached parents / user presets after system presets: checked by the full-set test (identical in every mode).
• Preset quarantine: user-preset path only, single-threaded, untouched.
• Parallel agents: 5B-5 (Orca port 5B-5: plates, instances and workflow UI (7 PRs) #679) also edits GUI_App.cpp and tests/libslic3r/CMakeLists.txt; git merge-tree against it is clean. 8-1, OD5 and OD6 edit option lists in PresetBundle.cpp / Preset.cpp, different hunks (OD5 merge-tree clean). 4B-3 (cache design) would build on the overrides added here.

Tests

• Build: edgebuild job 484 (vm) at ee380e3. Targets: libslic3r_tests, slic3rutils_tests, the app, and an install to C:/Dev/EdgeSlicerTest/pkg_4b2. The test filter was ~[NotWorking]~[.],[PresetLoadBenchmark], so that run included the full-set test, which was not yet hidden.
• libslic3r_tests: 2351 cases, 2348 passed, 3 failed as expected (402815 assertions). The 3 are the known [!shouldfail] cases in test_mixed_filament and test_wipe_tower_estimate. slic3rutils_tests: all 1403 cases passed (19274 assertions).
• The full-set determinism test passed: sequential == parallel cold == parallel warm == parallel, and parallel == parallel.
• The last commit only adds the [.] tag to the full-set test and was not rebuilt.

Timings (system + user preset load, load_presets(), full shipped set, VM, 5 runs each)

These are without the GUI, from the hidden benchmark.

Mode	median	min	max
sequential, no cache (upstream before OrcaSlicer#13667)	8068 ms	8042	8237
parallel, ORCA_PRESET_CACHE=0	3008 ms	2738	3118
parallel, cache on, cold (parses and writes the cache)	3311 ms	2740	3600
parallel, cache on, warm (= today's default on unchanged profiles)	3550 ms	3505	3575

Branch vs main: the load code is the same except that the two env flags are now read through a function, so main's numbers are the same rows. The win that OrcaSlicer#13667 promises (8.1 s to 3.0 s, -63%) has been on main since 2026-09-13. The warm cache costs about +0.5 s over cache-off; that's 4B-3a's decision. 4B-1's startup log put preset_bundle->load_presets at 3.2-3.8 s on a 16-model set, which matches.

G-code gate

Baseline _baseline 1009229 (CLI, scratch datadir). Main past it has only #650, post-processing scripts. Candidate = test copy C:/Dev/EdgeSlicerTest/pkg_4b2, edgebuild --gate, estimates ignored.

case                result             lines    -del    +add  hunks   cfg  est
x1c_cube            IDENTICAL          12321       0       0      0     0  same
x1c_benchy          IDENTICAL          98486       0       0      0     0  same
u1_cube             IDENTICAL          11098       0       0      0     0  same
u1_benchy           IDENTICAL         128144       0       0      0     0  same
u1_2col             IDENTICAL         149526       0       0      0     0  same
u1_t4               IDENTICAL         107372       0       0      0     0  same
h2d_benchy_2n       IDENTICAL         153939       0       0      0     0  same
h2c_benchy_2n       IDENTICAL         158235       0       0      0     0  same
voron_cube          IDENTICAL          12295       0       0      0     0  same
voron_benchy        IDENTICAL         176776       0       0      0     0  same
GATE_RESULT identical=10 config_only=0 diff=0 expected=0 failed=0 missing=0

No G-code change was expected either: the only non-test code is the splash (GUI) and the reading of two env flags.

Owner hand test

Test copy: C:/Dev/EdgeSlicerTest/pkg_4b2

1. Cold start. Close EdgeSlicer, start the test copy, and time it with a stopwatch or the log timestamps ("begin to show the splash screen" to "main frame firstly shown") against a main install. Expect the same time; this PR doesn't change loading. On the splash, the slicing animation keeps moving. The line under the logo goes "Loading configuration..." > "Loading printer & filament profiles..." > "Creating main window..." > "Loading current presets..." > "Showing main window...". A thin red bar under it fills in steps and never goes back, and the splash closes when the window appears.
2. Click the splash during startup: it closes and startup carries on (no crash).
3. Printer switching. Switch X1C > U1 > H2D > X1C: presets, filament slots and flow-variant combos look the same as on main.
4. Project presets. Open a project with project presets: they're listed and selected. Then File > New Project: they're gone.
5. User presets, including detached ones. Your own filament/process presets are listed. A detached user preset and any preset inheriting it both appear with their own values.
   No printer needed. Mac: not needed (no platform-specific code; the OSX wxYield in SetText is kept).