The spread was measured on opaque colours only, and the issue keeps
today's behaviour for partly transparent ones, in the match and in the
anti-alias fringe.
On 8-bit RGB a pixel now matches the clicked colour when the spread of
{0, dR, dG, dB} is within the Tolerance, so opposite changes in two
channels count double and green and blue stay at Tolerance 255. The
anti-alias fringe uses the same distance. The Magic Wand, the Paint
Bucket and other modes and depths keep the per-channel limit.
Instead of turning Anti-alias off, the tests keep the default setup and
expect the reference application's measured results (#11776, #11779):
the selection ends at 381, Subtract at 219, Intersect runs 219-381.
Suggested by hectorviov on #12399.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
#12146 gives near-matching neighbours a soft fringe when Anti-alias is on.
The #11848 tests left Anti-alias at its default and expected exact column
bounds, so main's engine tests failed (381 instead of 380).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On a transparency-locked layer with a colour-channel target, the live
preview mixed the dabs back with the lock (straight mix) while the commit
mixes premultiplied without it, so at 32 bits partly transparent pixels
rounded differently in the preview. The channel branch now builds the
LiveDab unlocked and pushes the points afterwards, as the RGB branch does.
The test now runs at 8, 16 and 32 bits over a half-transparent strip at
half opacity; it failed at 32 bits before the fix.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The review on #11853 noted that the tests checked only that a selection
row's label did not start with "Count", never tested `auto` choosing a
selection, never tested a ruler row, and read only the first row. Now
each record must give exactly one row, the selection and ruler rows must
be "Measurement N", `auto` with a selection must stay "Measurement N",
and the shared counter is checked across all three sources.
Co-Authored-By: DuBearte <DuBearte@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Cause: is_hard_rect read selection coverage over the selection's own content bounds in one buffer. After a transform those bounds can reach far past the canvas, while the patch-size guard only measures the on-canvas part, so a huge mostly-off-canvas selection could stall or exhaust memory.
Fix: is_hard_rect now calls is_hard_rect_within with MAX_PIXELS (the patch's own cap); bounds over the cap are not a hard rectangle and the selection stays in place. Otherwise coverage is scanned in 64-row bands, stopping at the first pixel below 1.0.
Verified: new test covers a limit below and at the pixel count and a hole in the third band; patch (25) and retouch (126) lib tests pass, clippy -D warnings and fmt clean.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Model-Effort: Opus 5.5 drove and briefed; Sonnet 5.5 wrote at medium
The engine turned every "mode": "clear" stroke on the Background into
Normal, so a held ~ painted the foreground there instead of erasing to
the background colour as on main. The options bar already turns a chosen
Clear off at the press, so the engine conversion was not needed: drop it
and keep the one check in clear_off_on_background.
A Background with locked pixels refuses the stroke, so a chosen Clear now
stays chosen there. The "Tool state machine" doc line is back on
tool_event.
Co-Authored-By: DuBearte <DuBearte@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Measured in the reference application:
- A group writes its box (the canvas) and z-index, then the rule of each
layer in it, top first, with no blank line between rules. A group's end
takes a place in the stack, so z-index counts it.
- A Solid Color fill layer without a shape greys Copy CSS out; so does a
selection with no layer that has CSS.
- Point type is placed by its line box: left is the pixel left less the
ink's offset from the origin; top centres the line box on the pixels, or
on half the cap height above the baseline when the text has a
descender. Left and top are cut after three decimals, not rounded.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Cause: the menu bar, the command labels and the selection tools' canvas menu spelled the two commands "Layer via Copy" and "Layer via Cut", while the reference application (and PhotoCraft's own History names and ATN event table) use a capital V.
Fix: renamed the labels in crates/ui-egui/src/menu_catalog.rs, the two spec! labels in crates/engine/src/cmd/edit_cmds.rs, and the rows in crates/ui-egui/src/canvas/canvas_tool_menu.rs with its test assertions. In the 33 locale files crates/ui-egui/src/i18n/*.tsv that carry the keys, only the key column of those two rows changed (66 rows), so every translation is kept and no language falls back to English. New tests are in crates/ui-egui/src/layer_via_label_tests.rs, registered with one mod line in lib.rs like the other *_tests.rs files.
Verified: the 5 new tests pass; canvas_tool_menu 51 passed; engine edit_cmds 112 passed; i18n 66 passed and 8 live_tests failed with the known local font panic in epaint fonts.rs (font-dependent, green on CI); cargo xtask parity left docs/parity-checklist.md unchanged; clippy -D warnings and fmt are clean. TSV check against HEAD: 34 files, same row count in each, exactly 66 rows differ (2 per file in 33 files, Hindi has none), only the key column of those rows differs, the translation column is byte-identical, no old key remains, and no CR bytes in any touched file.
Not changed: the translations' text, the History names and the ATN event table (already the capital V), docs/ui-parity.md and the other parity docs.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Model-Effort: Opus 5.5 drove and briefed; Sonnet 5.5 wrote at medium
Since the embedded-LUT params were added, the generated Color Lookup dialog
filled `fileName` with 0 (the dialog has no kind for `text?`), `data` with ""
and `lutFormat` with "cube". Every OK failed, and a .3dl or .look file from
the LUT folder (#9754) would have been parsed as .cube. `fileName` is now
plain text, `lutFormat` gets an `auto` default (detect from the file), and
an empty `data` / `fileName` counts as not given, like an empty `file`.
Also puts the Lens Blur test's doc comment back on its test (a merge had
moved it onto the LUT folder test).
Co-Authored-By: Luis-0508 <43929895+Luis-0508@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Conflict in retouch_cmds/patch.rs with #12255 (one History name, "Patch Selection", for every mode): kept the Destination branch that moves a hard rectangular selection with the patch and gave it HISTORY_NAME, and restored Plan's destination flag, which #12255 had removed as unused. patch tests 24 passed (including #12255's history-name test), clippy and fmt clean.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Model-Effort: Opus 5.5 drove and resolved
Cause: is_hard_rect required the selection's bounds to equal plan.area, which is clipped to the canvas. A hard rectangle reaching past the canvas (e.g. moved there with select.transformSelection, which keeps the off-canvas pixels) has bounds larger than the area, so the pixels were patched but the selection was left behind.
Fix: treat the selection as a hard rectangle when nothing is selected outside its own bounds and every pixel inside those bounds is fully selected, then translate the whole selection by the drag offset. Feathered, non-rectangular and holed selections still stay in place.
Verified: new test destination_moves_a_hard_rectangle_that_extends_past_the_canvas failed before the change (selection stayed at -20,100..40,140 while the pixels changed) and passes after, including one-undo restore of pixels and selection. cargo test -p photocraft-engine --lib patch (19 passed) and retouch (120 passed), clippy -D warnings and fmt clean.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Model-Effort: Opus 5.5 drove and briefed; Sonnet 5.5 wrote at medium
PR #11848 built the full-area selection mask for every Subtract and
Intersect click, though the global match (Contiguous off) ignores it.
Build it only when the fill is contiguous.
Tests for the review: Contiguous off keeps the global match, Subtract and
Intersect without a selection select nothing, the row check fails on a hole
in the selected run, and the fill's half-selected threshold (0.49 vs 0.5)
at the seed and at the fill's edge.
Co-Authored-By: DuBearte <130190089+DuBearte@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Destination branch of the Patch command no longer clones the whole document: it takes only the selection before the target surface borrows the document, and computes the selection's content bounds once, passing them to `is_hard_rect` and reusing them for the translate. Behaviour is unchanged.
Tests are split and extended: Destination with a feathered, ellipse, lasso and holed-rectangle selection leaves the selection in place while the patch still lands; Destination with no selection fails, adds no history entry and a following undo reverts the previous step, not the setup; redo restores both the patched pixels and the moved selection; Source and Content-Aware are separate tests, the latter asserting a pixel change; the main test also asserts source pixels unchanged, the destination carrying the source texture, the moved selection still hard-edged, and the returned damage and offset. A rectangle larger than the patched area and a partly covering selection without a hole cannot be built, because the patched area is the selection's own content bounds; an inverted selection makes the area the whole canvas, which Patch refuses.
Verified: `cargo test -p photocraft-engine --lib patch` 18 passed; `--lib retouch` 119 passed; clippy `--all-targets -D warnings` clean; `cargo fmt -p photocraft-engine` clean.
Not changed: the coverage pass in `is_hard_rect` (one pass over the patched area per command, the same order of work as the patch itself).
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Model-Effort: Opus 5.5 drove and briefed; Sonnet 5.5 wrote at medium
trim_name drops white space, ZWNJ/ZWJ and the blank characters (Braille blank, Hangul fillers,
invisible operators, tag and other default-ignorable blanks) from both ends. A variation or
Mongolian free variation selector or CGJ that belongs to a kept character stays, and so do Hangul
filler jamo that complete a syllable. History names use the same trim as plug-in names.