Skylar's review of 886a0b3:
1. tc/tid go locale by locale: in each locale the context or @id row,
then the plain row, so pt-br's plain row beats pt's context row.
2. pt (and pt-* other than pt-br) uses pt-PT plural rules: 0 is plural.
3. A script subtag (sr-Latn-RS, sr_RS@latin, be_BY@latin) never picks a
code of another script; sr-Cyrl-RS still picks sr.
4. A row defined twice is reported and the first kept, as in the checker.
5. The pseudo locale is debug-only via from_code, get marks too, and
looked_up() lists what was looked up for an app's audit.
6. intl_pluralrules pinned to =7.0.2; plural_form_counts copied from the
checker, plus pt.
7. I18n::new takes the app's offered codes: languages(), from_code,
match_tag and system_lang only return those; others stay fallbacks.
Smaller: en/pseudo/repeated codes are skipped and reported,
system_lang is cached, zh-hant-hk falls back to zh-hant, rtl adds ckb,
dv, ps, sd, ug and yi.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Review fixes (CodeRabbit): a glibc @latin/@cyrillic modifier becomes the
script subtag and only codes with that script are tried. sys-locale
drops the modifier, so on desktop Unix system_lang reads LANGUAGE,
LC_ALL, LC_MESSAGES and LANG raw, in its order. README: tl! qualifies
CURRENT with $crate; get follows the locale chain but never the plain
row or English.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Final review fixes: a European Portuguese locale wins over pt-br when both
are embedded; sr-latn no longer falls back to Cyrillic sr; get() looks up
one context with no plain or English fallback (key-cap names).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01F9fM3mv1PUyKpk5rdwUQrW
resolve_absolute_refuses_results_that_are_not_absolute resolved links against a
document at /home/u/docs/a.doc outside its cfg!(windows) branch. On Windows that
path is rooted but has no drive, so it is not absolute, resolve_absolute refuses
the result as documented, and the unwrap panicked. The Windows arm already covers
the same three cases with a drive. Found by the new docpath CI job on its first run.
A test is not instant and the address field stays editable while it runs,
so its answer can arrive after the address has changed. `set_models` then
recorded the models against whatever the field held by the time it
returned, filing one endpoint's models under another — the mislabelling
`models_fit` exists to prevent, reached the asynchronous way rather than
the synchronous ones already closed.
`set_models` now takes the address the test was started against and
returns whether the models were taken, so an answer that outlived its
address is dropped and the caller can say the result is out of date. The
parameter replaces the old signature rather than joining it: there was no
second legitimate caller, and leaving an address-free version beside it
would have left the racy call the shorter, easier one to reach — against
this module's rule that the rules live once so two screens cannot
disagree.
Also fixes a pre-existing mismatch found while testing this:
`from_connection` recorded the loaded address untrimmed while `prepare`
compares it trimmed, so a stored endpoint with stray whitespace dropped
its models on the first save.
Model provenance across loaded and preset endpoints and repeated endpoint edits, so models from one endpoint are not relabelled as another's, and eligibility resolved in both directions: a model must accept every input the tool sends, and the tool must supply every input the model requires (new ResolveError::InputRequired).
All 18 checks pass on a40245bd, including craft-ai-core fmt, clippy -D warnings, tests and the wasm32 check on Ubuntu, Windows and macOS.
Note for callers: after a tracked endpoint change, use set_models for newly discovered models; direct assignment on fresh custom drafts and unchanged saved endpoints still works. The asynchronous in-flight probe race is separate and not addressed here.
resolve_absolute refuses a result that is not absolute on the running
platform, such as a relative link in a document that has no path yet.
It is the function to call before opening a file.
The eager tuple computed o + 8 even when the first read had failed,
overflowing a 32-bit usize at offsets near u32::MAX (n = 2 with
size = u32::MAX - 20 reaches it: o fits, o + 8 does not). The record is
now bounds-checked in one slice and the fields read inside it, the same
complete-record pattern the profile reader uses elsewhere.
Consent expiry is now per kind of data: renewing text no longer renews an expired image permission. A connection edit keeps models found by a test at the new address, and drops only the ones saved for the old one.
parse_curve computed 12 + n * 2 with n straight from the file and only
checked the 1 << 20 bound afterwards, inside the table arm — on a 32-bit
(wasm32) target, which craft-cms ships branches for, a trC curve count of
0xFFFFFFFF overflows with overflow checks on and panics before the guard
could reject it. The mluc reader computed 16 + k * size unchecked, where
wrapping reads another record's bytes as the language, length and offset
(wrong text, no error) instead of stopping.
Hoist the curve-count guard above the arithmetic and compute the record
offset with checked_mul/checked_add, breaking out when it does not fit.
This ports the app-side fix accepted as PhotoCraft #4976 / PR #5227,
which was never carried into craft-cms. On a 64-bit host both sites are
benign (the arithmetic fits), so the two new tests pass before and after
there; they bite on a 32-bit/overflow-check build, which deserves a
wasm32 CI job — craft-cms has none yet.
Add opt-in shared PDF authoring with bounded font preparation, caller-owned metadata and settings, guarded serialization, and a backend that rejects poisoned drawing state.
Preserve the existing writer commits so immutable dependency pins remain reachable from main. Reader and application migrations remain separate.
PR #105.
A configured connection does not authorise an app by itself. An agreement now
covers one app, connection, destination and capability, for the kinds of data
it named, and may carry an expiry (unix seconds).
- Disclosure carries the capability; ConsentRecord gains optional capability
and expires_at.
- ConsentLedger::needs_consent takes the time (the crate has no clock);
grant_until sets an expiry, grant stays open-ended. The latest grant sets
the expiry for that capability only.
- Records saved before this have no capability and cover nothing, so the
person is asked once more.
Nine new tests; six deliberate bugs (expiry ignored, off by one, capability
ignored, legacy record covers everything, expiry not updated, timed grant made
permanent) were each caught.
Builtin::from_profile_name matches only a built-in's own description.
Builtin::stand_in_for adds the FOGRA39 approximations (Coated FOGRA39, ISO
Coated v2) for documents that name a profile without carrying its bytes.
ISO Coated v2 300% (ECI) has no stand-in: the built-in separates to up to
330 % total ink. The README example parses the document's ICC bytes first
and handles a missing stand-in explicitly.
The reference value for CMYK 0/41/90/4 in Coated FOGRA39 is 239,161,35.
The fitted profile gives 239,161,33 (ΔE00 0.16); the uncalibrated formula
DesignCraft used gives 245,144,25 (ΔE00 6.36). Six more reference values
(K50 grey, rich black, light tint, skin tone, C+Y green, near-neutral)
are within 0.7 ΔE00.
The Generic CMYK oracle only runs where macOS ships that profile. The
same comparison (CMYK -> sRGB, 8-bit, relative colorimetric without
BPC, same bytes for both engines) with the built-in Coated FOGRA39
runs everywhere. Measured: max 10, p99 5, mean 0.89 codes; the worst
pixel matches an independent float computation, not moxcms.
With a LUT destination, sRGB white came out as 0.3-0.5 % C/M/Y through
Transform::eval and precise float conversions: the legacy Lab encoding
puts L* 100 at 0.996 and a* = 0 at 0.50001, between the BToA grid
nodes, and fix_white only patched the device-link node, so only the
table paths were right.
Transform: for every intent but absolute colorimetric (and proofs
without paper simulation), a pixel equal to the source's device white
gives exactly the destination's device white. fix_white still sets the
device-link node; the paths that do not read that node exactly (the
exact pipeline, float interpolation, and 8-bit lookups when white is no
node) compare each pixel with white once. No other colour changes.
device_white(Lab) had a* = b* = 0.5; Lab white is 128/255 in the v4
encoding, which is no node of the 33-point grid, so Lab sources take
the per-pixel check there too.
Profile: synth::CmykParams::white_on_node maps no ink to exactly
L* 100, a* = b* = 0 in the AToB and gives the BToA tables input curves
that put that white on a node holding no ink. The Coated FOGRA39
profile is regenerated with it and separates PCS white to 2e-5 ink
(was 0.5 %), so the step into white is as smooth as its neighbours.
The default parameters and the Photocraft profile are unchanged; its
white still falls between nodes, so its last grey step into white is
up to about 0.4 % C larger than the one before, as with lcms's
white-node fix, and the test allows that for this profile.
FOGRA39 accuracy against 16a66e1: AToB against the patches unchanged
(mean 0.07, p95 0.18, max 0.53 ΔE00), reference sRGB values unchanged
(max 0.76), Lab -> CMYK -> Lab p95 1.04 -> 1.07, max 2.54 -> 2.55
(mean 0.24).
The crate denies panics, but asserts and slice indexing were reachable
from arguments: pixel strides, buffers that are not whole pixels or a
destination shorter than the source, single-colour slices shorter than
the channel count, CLUT and pipeline dimensions, and LUT sizes.
Every such check now returns CmsError::InvalidArgument. Buffer
conversions validate once per call, before any pixel is written, so the
per-pixel loops carry no extra checks; internal callers that already
guarantee lengths use crate-private unchecked evaluation.
Pipeline, Clut and Lut3d keep their fields private (read through
accessors) so a validated value cannot be changed into one that panics.
Pipeline::new checks matrix shapes and channel counts; Clut::new and
Clut::sample check grid, outputs and the table size limit.
API changes (unreleased crate):
- Transform::eval, eval_fast, convert_u8/u16/f32, apply,
convert_bytes_many and from_pipeline return Result.
- Pipeline::new, extend and eval, Clut::new, sample, eval,
eval_trilinear and eval_multilinear return Result.
- GamutCheck::delta_e and out_of_gamut, Lut3d::from_transform,
LutFile::identity and from_fn, synth::cmyk_profile and coated_cmyk
return Result.
- Pipeline, Clut and Lut3d fields are read through methods.
Add Builtin::CoatedFogra39 ("Coated FOGRA39 (fitted, craft-cms)",
profiles/craft-coated-fogra39.icc: AToB 11^4, BToA 33^3, 330% TAC,
medium GCR). It is fitted to Fogra's FOGRA39 characterisation data
(tests/data/FOGRA39.txt, unmodified, 1617 patches): a Yule-Nielsen
Neugebauer model with free overprints, per-ink tone curves and n, plus
a Gaussian-kernel ridge correction of the residuals. Against the
patches, relative colorimetric through the AToB: mean dE00 0.07,
p95 0.18, max 0.53. The BToA tables invert the AToB table
(synth::cmyk_profile_from_model); the Photocraft coated profile and its
synth defaults are unchanged.
Builtin::from_profile_name maps the names documents carry for
FOGRA39-based profiles (Coated FOGRA39, ISO Coated v2 (ECI)) to the new
profile and every built-in by its description. Adds math::delta_e2000.
Refs #36
Martin BallandClaude Opus 5.5
skylar taylor-barrick
aiongg
midasmatterhorn
Jeremy
skylartaylor
skylartaylor
