* Use chrono for UTC/calendar conversion instead of hand-rolled copies
Replaces the hand-rolled Howard Hinnant civil/days-from-civil copies and the
leap-year/days-in-month table with chrono as a workspace dependency:
- apps/lightcraft/src/logging.rs: timestamp (log line time)
- crates/catalog/src/dates.rs: civil + weekday
- crates/catalog/src/stacks.rs: iso_seconds (date -> seconds)
- crates/engine/src/cmd/face_recognize.rs: capture_secs (native only)
- crates/meta/src/datetime.rs: DateTime::unix_seconds and the valid() day check
Output formats and semantics are unchanged (pre-epoch clamp, day rollover that
normalize_iso relies on, leap-second tolerance, EXIF/XMP parsing). chrono uses
only std, so wasm stays safe. AGENTS.md/CLAUDE.md now say time goes through chrono.
Follow-up to storytold/cadcraft#1785
* Never panic on a clock outside chrono's range; keep 5-digit years unprefixed
chrono's From<SystemTime> unwraps, so a clock past year 262143 panicked
inside the logger. timestamp now uses the fallible DateTime::from_timestamp
and reads such a clock as the epoch, keeping the old pre-epoch clamp.
timestamp and catalog's civil write years past 9999 with {:04} like the old
{y:04} did, because %Y would add a '+'. Tests cover the out-of-range and
5-digit-year cases.
* Don't build an unrepresentable SystemTime in the far-future test
UNIX_EPOCH + Duration::from_secs(u64::MAX / 2) overflows Windows'
SystemTime and panicked before timestamp() ran, failing the Windows
test job. Build it with checked_add and only assert where it exists.
* time: preserve old out-of-range date conversion (no epoch fallback)
`DateTime::from_timestamp(...).unwrap_or_default()` silently reset any instant
past chrono's year range to 1970-01-01, changing output the old hand-rolled
Howard-Hinnant code converted exactly (e.g. 10_000_000_000_000 s was
318857-05-20 17:46:40, not the epoch). Restore the old conversion domain with a
cycle-based helper that routes the calendar through chrono (the calendar
repeats every 400 years = 146_097 days, so month/day/time come from chrono and
only the year is carried in i128, keeping every i64 exact).
- apps/lightcraft/src/logging.rs: `timestamp` via `utc_parts`; keep the
pre-1970 epoch clamp; u64 secs past i64::MAX saturate, not epoch.
- crates/catalog/src/dates.rs: `civil` cycle-based (the callers stay inside
0000-9999, but a far clock no longer resets to the epoch).
Pinned old outputs in tests:
logging u64::MAX/2 s -> 292277026596-12-04T15:30:07.000Z (was epoch),
10_000_000_000_000 s -> 318857-05-20T17:46:40.000Z
civil 10_000_000_000_000 -> 318857-05-20T17:46:40,
-10_000_000_000_000 -> -314918-08-13T06:13:20, i64 extremes
* Don't build a year-318857 SystemTime unchecked in the logging test
Windows' SystemTime only reaches about year 30828, so
UNIX_EPOCH + Duration::from_secs(10_000_000_000_000) panicked before
timestamp() ran and failed the Windows test job. Build it with
checked_add and assert where the platform can hold it. The year-10000
case fits everywhere and stays unconditional.
* Name the 400-year range extension as the one exception to the time rule
The rule said never to hand-roll calendar math, but the range fix
(shifting by whole 400-year Gregorian cycles so dates past chrono's year
range keep their old output) is cycle arithmetic, so the rule and the code
disagreed. The rule now names that exception: the library still computes
month, day and time, and only whole cycles are carried in the year.
* Say the library computes month and day; time-of-day is plain arithmetic
Narrows the range-extension wording: the date library computes the
calendar fields, and splitting seconds into hours, minutes and seconds
is ordinary arithmetic, which is what the code does.
Several commands read their photo parameters leniently. An `ids` that
wasn't a list of whole numbers (`"ids": 3`, `["3"]`) fell back to the
selection or the active photo, and an `id` that wasn't a whole number
(`"id": "5"`) fell back to the active photo, so a call meant for one
photo acted on another:
- ids: album.addPhotos / removePhotos / toggleTarget, library.select,
develop.set, denoise.queue, faces.index, mask.applyEffect (mask ids)
and app.export;
- id: photo.relink (relinked the active photo to the file),
photo.removeRegion / setRegion, faces.setName, stack.setTop / moveUp /
moveDown / split, library.findSimilar, the photo queries,
denoise.cancel (cancelled every photo) / status and ui.render.
They now refuse such values with an error, as Session::targets already
does for the other commands (shared cmd::photo_ids / photo_id). A
missing or null value still means the selection or the active photo.
A wrongly typed id or ids fell back to the selection, or kept only the valid part of the list. Session::targets now returns an error for those, and callers pass it on.
Activations stay on the device between layers (net/program.rs lowers the graph; every
operator is a WGSL kernel, one submission and read-back per run); faces of a photo go
in batches of up to 16; Settings > Faces shows each model's device and a CPU-only
switch (faces.gpu); the scan's pace limits GPU use; the GPU check is an activity row.
YuNet 58 -> 21 ms, SFace 40 -> 9.6 ms, AuraFace 295 -> 71 ms (heavily loaded machine).
Sources: own design.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
An accelerator hook in the face interpreter takes Conv nodes; lightcraft-gpu implements it with a WGSL
implicit-GEMM convolution (any stride, padding, grouping; split-K for small maps; a direct kernel for depthwise
layers), each model's weights resident on a device of its own. After loading, detection and recognition check
the GPU in the background (agreement within 1e-3 and faster than the CPU) and otherwise stay on the CPU, which
remains the reference; the first GPU error drops it. Sources: own design.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
While photos are left to search for faces the stack shows "Finding
faces" with how many are done (no cross: the scan is switched off in
Settings). The engine keeps the scan's peak, so Settings and the stack
agree. Also translates the "Downloading model" label.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Keep current upstream UI behavior and the later detector contract checks. Validate candidate models before replacing working weights and records, with a synthetic regression that failed before the fix. Remove the unrelated GPU equivalence-test relaxation. All seven quality gates pass with the pinned hosted fonts.
Generated with Codex.
aiongg