* feat(engine): relative leading, baseline shift and tracking steps in type.setStyle
`leadingBy`, `baselineShiftBy` and `trackingBy` move every character in the
range from its own value, so mixed values keep their differences. Auto
leading starts from the value layout uses (the paragraph's auto-leading
percentage of the character's size), results clamp to Photoshop's ranges,
and non-finite or oversized steps are refused as bad params. `leadingBy`
follows `metricsAsShown` like `leading`.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* feat(ui): Photoshop's type formatting keys while editing text
While a Type tool edit is active, Shift+Cmd+K/U/B/I and Shift+Cmd+/ toggle
All Caps, Underline, Faux Bold, Faux Italic and Strikethrough on the
selection (off when every selected character has it, else on);
Alt+Up/Down step leading by 2 pt and Shift+Alt+Up/Down baseline shift by
2 pt (10 with Cmd); Alt+Left/Right over a selection track it by 20 (100
with Cmd), a collapsed caret still kerning; Cmd+Alt+Shift+H toggles
Hyphenate for the paragraphs touched. Each press is one type.setStyle
history step, and the keys no longer fall through to Color Settings,
Desaturate, Auto Color or Inverse while typing; outside an edit those
shortcuts work as before.
A collapsed caret in non-empty text changes nothing: Photoshop styles the
next characters typed, which the document model can't hold apart from the
text. In empty text the style becomes the typing style.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
* Camera Raw Filter: save and load settings as .xmp (#4274) [WIP]
* Camera Raw Filter: Save Settings… and Load Settings… (.xmp) (#4274)
Adds Save Settings… and Load Settings… to the Camera Raw Filter dialog's
footer. They write and read Adobe Camera Raw settings files: the crs:
namespace on x:xmpmeta/rdf:RDF/rdf:Description, as attributes or elements.
photocraft_io::camera_raw_xmp maps every control the dialog has:
- Basic (Exposure2012 … Saturation, IncrementalTemperature/Tint)
- the parametric and point tone curves (ToneCurvePV2012[Red|Green|Blue])
- the HSL mixer
- Color Grading (including the SplitToning* names for the shadow and
highlight wheels)
- sharpening and noise reduction
- grain
- post-crop vignetting
Loading applies only what the file holds over the current settings, as
Camera Raw does. Values are clamped to the slider ranges. Unsupported
settings (lens, calibration, geometry, defringe) are ignored, and the
status bar counts them when they aren't neutral. Malformed files fail
with a message and leave the settings unchanged. The control channel
takes {"ui":{"saveSettings"|"loadSettings":true}}.
Translations are added for all 17 catalogs.
Fixes#4274
* Add Turkish and Hungarian rows for the Camera Raw settings strings
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Brandon Thomas <echelon@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Draw the active layer's storage tiles over the canvas: a faint
TILE_SIZE grid over the document, every stored tile filled and
labelled tx,ty, tiles with identical pixels in one colour per group
and unique tiles in blue. It shows how sparse a layer is, which tiles
copy-on-write shares, and which ones .pcraft saves once.
It has no menu item and isn't a View > Show item. The control channel
and MCP turn it on with ui.set {"tileGrid": true}; ui.inspect reports
it as view.tile_grid, and View > Extras doesn't hide it.
Grouping hashes each distinct tile allocation once, in parallel (or
serially while jobs hold the pool, #1554), cached per document
revision and active layer. A 36 MP layer of all-different tiles groups
in 3.9 ms (8-bit) and 7.5 ms (16-bit) on 24 cores in release; the
ignored tile_grouping_36mp test reproduces this.
Also documents tileGrid in docs/control-protocol.md and
docs/development.md, and adds my contributors/people.toml entry.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
A large scale (e.g. 200x on a thin column) made the warp fill every tile of
a result up to 2^21 px on a side, enough to exhaust memory. Free Transform
now prices each warp before building anything and refuses with an
allocation-style error ("...; the document was not changed") once a result
would exceed MAX_TRANSFORM_PIXELS (2^30, the same bound as Puppet Warp's
MAX_WARP_PIXELS). Only the result's size counts, so moving content far off
the canvas still works, and nothing is clipped.
The check covers the layer, a smart object's cache and filter mask, linked
masks, a lone mask/channel target and the selection outline, and runs over
the whole layer tree before any layer is changed. Transform Selection gets
the same check. The transform box now stays open when a commit is refused.
Part of #3773
Co-authored-by: Fabio Lussi <198260595+itrycodinganditsbad@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* Type: faux bold keeps antialiased edges at small sizes (#4130)
Faux bold drew each glyph at 8 offsets into one signed-area coverage buffer, so edge pixels from the copies summed and saturated to full coverage (hard, pixelated stems, about 2.2x the ink at 12 px). Each offset now gets its own buffer, merged with max.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Type: faux bold ring radius tuned so stems keep their weight (#4130)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
layer.layerStyle.gradientOverlay, layer.layerStyle.stroke (gradient) and
paint.gradient take "method" (classic|perceptual|linear|smooth) and
"smoothness" (0..100, default 100). The stops are baked with the same
model PSD import uses (io::gradient_bake, now public as bake_gradient and
GradientMethod), so the compositors' linear interpolation reproduces
Photoshop's curve. Without a method the stops stay as they were.
Measured in Photoshop 27.11 (max level error vs main -> this change):
Gradient Overlay linear/reflected styles, any angle/scale/reverse 12 -> 0-1,
radial/diamond 16 -> 4-5; gradient Stroke 12 -> 2-4; Gradient tool
black -> white Classic 13-14 -> 2-3; red -> blue Classic 12 -> 1,
Perceptual 83 -> 2, Linear 64 -> 4, Smooth 83 -> 3. A 3-stop Classic
gradient improves only 14 -> 10 (the multi-stop smoothness spline is a
separate follow-up).
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* engine: remember the last Save for Web settings in the preferences
* engine: test that a complete export remembers its settings for the next launch
* ui: seed the Save for Web dialog from the remembered preferences
* ui: Export As opens with the settings it was used with last
* scorecard: FILE-215-18 done, the export dialogs remember their settings
* docs: scorecard, File compatibility 7 done of 27
* Regenerate the scorecard
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Brandon Thomas <echelon@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Clippy 1.96 flags the `retain` predicate that strips stale content blocks
from pixel layers on PSD export (`clippy::nonminimal_bool`), which breaks
`cargo clippy -- -D warnings` on current stable.
Rewrite it as a single negated disjunction. The four terms, their order and
the short-circuiting are unchanged, so the same blocks are kept and dropped.
CMYK samples are ink (1 = full ink), but both filters assumed light
polarity. Find Edges wrote 1 - edge, so flat CMYK areas became 100% ink
(black) instead of white; Solarize inverted channels above 50%, i.e. the
shadows. Both now use fxutil::subtractive(ctx), like Trace Contour: Find
Edges writes the edge strength as ink, Solarize inverts channels below
50% ink. Alpha stays untouched. RGB, Grayscale, Lab and other modes are
unchanged.
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Image > Mode > Bitmap (halftone) passed its angle through unchecked: 1e39
became inf as f32, the screen rotation turned it into NaN and an almost
white bitmap was recorded in history; 500 was accepted although the
params doc (and Photoshop's Halftone Screen dialog) says -180..180.
The angle must now be a finite number in -180..=180 and the frequency a
finite number (still clamped to 1..999 lines/inch); otherwise the command
returns BadParams before touching the document.
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
file.exrParts lists a part without a name attribute as "part <index>", but
file.openExrParts only matched parts whose name attribute equalled the
requested name, so "part/part 0" failed with "no flat part named part 0".
One helper now gives the display name for the listing, the lookup and the
layer name; a named part still wins over an unnamed one with the same
fallback name.
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Each Cryptomatte RGBA set stores two (ID, coverage) ranks, (red, green)
and (blue, alpha), but only red/green were read, so every second rank
was lost and mattes got holes at edges and overlaps. Suffixes were also
matched case-sensitively, missing Blender's .R/.G/.B/.A.
Stream channels now carry a rank (0 red/green, 1 blue/alpha), suffixes
match in any case, and decode stores each channel at slot*2+rank
(bounds-checked), still sorted by coverage. The allocation check counts
four channels per set.
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
rounded_rect took the absolute width/height without moving the origin, so
shape.create {"kind":"rect","rect":[50,10,-20,30]} covered x 50..70 while an
ellipse or polygon with the same rect covered 30..50. Normalise (x, w) and
(y, h) for negative sizes instead. The Shape tool drag already normalises its
rect, so interactive drawing is unchanged.
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
layer.vectorMask.edit stored `feather` as f.max(0.0) as f32, so 1e39 became
inf (hard edge, null in info, .pcraft save refused) and any huge value was
accepted. It now takes the same finite 0..=1000 px range and error as
layer.layerMask.edit, checked before the edit so a bad value leaves the mask
and history untouched. The params doc says so.
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The font menu's preview samples, family list and localized names were
cached on fonts::generation(), which any FontDb moves, and the previews
also on the served catalog's generation. Other engines registering fonts
(tests running in parallel) re-rendered every visible sample each pass,
and families covering Arabic asked for a repaint, so Harness::run never
settled: options_font_family_preserves_peer_faces_and_batches_layer_flags
and the font picker tests failed intermittently with "exceeded max_steps".
- photocraft-text: fonts::shared_generation(), moved only by the shared
engine's FontDb (FontDb::mark_shared from new_shared).
- ui-egui: families(), localized_families() and the preview cache key on
it; previews no longer key on the served catalog (a served family's
sample appears when its fonts register, which moves the generation).
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The warning triangle on a comp row sent layerComp.updateWarnings
{"clear":true}, which dropped the stale states of every comp. The
command now takes an optional "comp" (id or name) that limits both the
report and the clear to that comp; an unknown comp is an error and
changes nothing. Without "comp" it still acts on every comp (Clear All
Layer Comp Warnings). The row icon sends its own comp id.
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
32-bit samples can lie outside 0..1. The RGB readout clamped them, but
device_cmyk did not, so a pixel at 1.5 read K:-50% and negative values
gave C/M/Y above 100%. Clamp each channel to 0..1 (NaN reads as 0) before
the conversion; this fixes both the Color Sampler and the Info panel.
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
#3132 refused type.edit on a Lock All layer, but type.insertText,
edit.checkSpelling "change", Paste Lorem Ipsum (and the other
type_extra_cmds::with_text edits) and Find and Replace Text "change"
on a layer locked after the find still changed its text and recorded
history. The check is now one pub(crate) helper,
type_cmds::check_text_unlocked, called by every one of these paths
with the same "...is locked" error as type.edit.
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
* feat(ui): add GRD import option to gradient preset picker
* Translate the GRD import strings into Vietnamese, Turkish and Hungarian
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Brandon Thomas <echelon@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* WIP: Object Selection hover preview (#4087)
* Object Selection: Object Finder hover preview and click-to-select (#4087)
Hovering with Object Selection highlights the object a click selects,
computed off the UI thread after a short rest and cached per document
pixels; a click commits it via select.object {point} honouring
add/subtract/intersect; dragging keeps the rectangle mode. Adds an
'Object Finder' options-bar toggle (default on, persisted) with
translations in every catalog.
---------
Co-authored-by: Brandon Thomas <echelon@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* Layers: Fill and the Paint Bucket edit an adjustment layer's mask when its icon is targeted (#4298)
An adjustment layer has no pixels, but once its icon thumbnail was clicked (or it was
reached with Alt+[ / Alt+]) the Layers panel no longer targeted its mask, so Edit > Fill
(Shift+F5) and the fill keys were greyed out, and the Paint Bucket and brushes failed with
"not a pixel layer". As in Photoshop, painting and filling an adjustment layer now always
edit its mask; the panel's target (brackets, trash, Delete) is unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Keep the web read_dropped doc comment on its function
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
* fix(ui): keep brush settings section lock icons permanently visible (#4278)
* test(ui): verify brush settings locks stay visible and toggle on click (#4278)
* Format the lock-icon test
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Brandon Thomas <echelon@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(ui): add Brushes button at top of Brush Settings control list (#4265)
* test(ui): verify Brushes button at top of Brush Settings opens Brushes panel (#4265)
* fix(ui): use a linear gradient ramp glyph for the Gradient tool icon (#4172)
* test(ui): verify Gradient tool icon rasterizes a left-to-right ramp (#4172)
* fix(ui): make clickable tab and radio labels non-selectable with pointing hand cursor (#3972)
* test(ui): verify New Document preset tabs show pointing hand cursor instead of text caret (#3972)
The options bar's Maintain Aspect Ratio link only affected typed W/H values; dragging a
corner or edge handle still scaled proportionally, so unlinking seemed to do nothing. As
in Photoshop CC 2019 and later, the link now also sets the drag default: unlinked, corner
and edge drags stretch freely and Shift keeps the proportions. Linked (the default) and
the legacy preference are unchanged. The link state moves from egui temp data into
UiState (transform_unlinked), next to the transform session.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
In the Studio-layout themes (Studio, Studio Light, Breeze, Adwaita,
Solarized) Properties is a floating card over the canvas. The card
returned early unless the active layer was an adjustment or fill layer
or a mask was targeted, so for a pixel layer, the Background, type or a
group nothing was drawn, although Properties was on (the rail button
showed it). The Pro theme's docked panel shows all of those.
The card now renders the docked panel's body (properties_body), so both
show the same content: the Document properties for the Background or no
selection, layer, adjustment and mask properties. Its title bar, hide
button, position and drag are unchanged.
Tests: floating_properties_card_shows_document_properties_for_the_background
and floating_properties_card_shows_pixel_layer_properties (Studio and
Breeze Dark) failed before the fix.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The section list indented every row without an enable box by the box's
width, except the first, so Brush Tip Shape started in the checkbox
column while Smoothing, the other box-less row, lined up with the names.
Photoshop indents Brush Tip Shape too. Every name now starts in the
same column.
Test: section_names_line_up_whether_or_not_they_have_a_box checks the
painted x of every section name (Brush Tip Shape was 20 px left before).
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* Canvas: a tool drag past the edge auto-scrolls the view (#4183)
Dragging a marquee, lasso, stroke or crop off the visible part of a zoomed
image now scrolls toward the pointer, as in Photoshop, and stops at the
scroll limits. An axis where the image fits does not scroll, and an
overscrolled view is not pulled back.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VFAjTZQw6MTyhfpQqjAfoW
* Fix rotated canvas auto-scroll limits
---------
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: jonathnwxh <258340642+jonathanwxh-cell@users.noreply.github.com>
* Control: a timeout says whether the command ran (#3662)
A request still queued when its deadline passes is dropped before
dispatch, but it answered with the same text as the transport giving up
on a command that may still apply. The two now read differently: "timeout:
not run ... safe to re-send" and "timeout: no reply in 60s ... may still
apply", both keeping the `timeout` prefix machines match on.
* Control: a transport that gives up knows whether the command was dispatched (#3662)
The transport and the UI thread share the reply deadline, and the
transport's wait runs out first, so the UI thread's "not run" reply
almost never reached a client. A request now carries a dispatch flag
settled once between them: the UI thread claims it before running, the
transport abandons it when it stops waiting. The transport then replies
"not run" for a request that will never run, and "may still apply"
only for one that was dispatched; a request given up on is never run.
returned_non_graphics_errors_do_not_trigger_backend_recovery failed at
random in full `cargo test` runs: "gpu-error: ...", left: false,
right: true. Its gpu-error case drops the sentinel, which leaves the
marker file in place, and calls `Sentinel::begin` again for the same
directory. Tests in monitor_profile.rs spawn processes in the same
binary. A child process started while the marker is open gets a copy
of its file descriptor, and the marker stays locked through that copy
until the child's exec closes it. When `begin` runs in that window, it
returns `Previous::Busy` instead of the crashed marker, and the
assertion fails.
sentinel_detects_a_crash_and_clears_on_success has retried `begin` for
this since 3b7870be (#289). The retry moves into a test helper,
`begin_when_unlocked`, which also runs where
returned_non_graphics_errors_do_not_trigger_backend_recovery and
a_failed_cpu_start_does_not_pin_later_starts start again after
dropping or finishing their sentinel. In
a_failed_cpu_start_does_not_pin_later_starts the CPU plan finishes the
sentinel, which removes the marker, so the helper's first `begin`
finds the lock free; the helper covers that test's branch for a plan
that keeps the marker.
A new test, begin_when_unlocked_waits_for_a_briefly_held_lock, takes
the marker's lock through a second handle after the sentinel is
dropped, releases it from a thread after 50 ms, and checks that the
helper returns the crashed marker. It fails when the helper calls
`begin` only once.
* feat: resize Save for Web previews and added palette editing for Gif/PNG8
* Save for Web: add Turkish and Hungarian rows for the preview navigation hint
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Brandon Thomas <echelon@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
#4093 changed the Auto-Select default to Group, as in Photoshop. The two
layout_perf tests that click layers inside groups with Auto-Select on
expected the layer under the pointer, so they now select the group and
fail on main. They set the target to Layer explicitly, which is what
they test.
layer.layerStyle.globalLight cast its params straight to f32, so an angle
of 1e39 became inf and rem_euclid turned it into NaN, which was stored,
recorded as a history step and fed to every effect using global light.
Angle and altitude are now read as f64 and rejected with BadParams when
non-finite or beyond f32 range, before the edit. In-range wrapping and the
altitude clamp are unchanged.
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The writer capped names with encode_utf16().take(1024), which could end
on a lone high surrogate; it now uses io::utf16_units, like .aco/.ase,
so the cap falls on a character boundary.
The reader's "count left out the null" check looked at the last unit
kept, not the last unit read. For a name of 1024+ units the counted null
is read but not kept, so a following 0x0000 (an empty contour's point
count) was swallowed and the stream shifted. It now checks the last unit
actually read.
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
With a selection, Edit > Transform > Warp warped only the selected pixels
but the whole linked layer mask, so the unselected mask no longer lined up
with the unselected pixels. Warp now limits the linked mask to the selection
through transform_cmds::move_gray_selected, as Free Transform and
Perspective Warp already do (#3083); layers that move whole still take their
whole mask along. Adds a regression test driven by edit.transform.warp.
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Every video-layer edit ran `work_end = work_end.max(duration)`, so
Show Altered Video, Restore Frame, Interpret Footage and the other frame
edits reset a narrowed work area to the full duration. A new
`grow_duration` helper, used by every video edit and by New Video Layer
from File, grows the timeline and moves the work-area end only when it
reached the old end; a narrowed work area stays where the user put it.
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
paint.bucket and paint.gradient read `opacity` as f(p,"opacity",100)/100
with no range check, so 200 became a coverage of 2 and overshot the blend
(#808080 on white painted near-black). Both now use the Magic Eraser's
param reader: the value is clamped to 1..100, and a non-numeric opacity is
BadParams before any edit or history step.
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Root cause: the JSON-lines server's `index_of` read `index` with `as_u64`, so a
string, negative, fractional or boolean index became None, which means "the
active document". `doc.close {"index":"0"}` therefore closed the active
(possibly unsaved) document instead of failing. The MCP server types the param
as Option<usize> and already rejects these values.
Fix: `index_of` returns Result<Option<usize>, AutomationError>. A missing or
null index still means the active document; a non-negative integer is used;
anything else is a bad request ("`index` must be a non-negative integer").
All doc.* call sites propagate the error with `?`.
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Root cause: the layer row's reserved width and indicators, and the Studio
themes' lock button, treated a layer as locked only for transparency,
position or all. A "Lock image pixels" or artboard lock left the row
without its lock icon, and the Studio button drew "unlocked" although
clicking it unlocked the layer (simple_lock_toggle checked all five).
Fix: add Locks::any() in photocraft-doc and use it in all four places,
so every lock shows the icon, as in Photoshop. Tests cover each lock on
the row (reserved width and Indicator::Lock) and the Studio button state.
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Root cause: the options bar sent markerSize/labelSize as f32 JSON floats
(8.0) and count.setGroup read them with as_u64, which is None for floats,
so the command returned Ok and changed nothing.
Fix: count.setGroup accepts whole numbers sent as floats (UIs, agents and
actions send them), rejects fractions, negatives and non-numbers with
BadParams, and clamps in u64/f64 so huge values no longer wrap through
`as u32`. The options bar now sends integers.
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Root cause: run_on_layer in distort_cmds.rs added the smart filter and
returned for a smart object before it called check_locks, so a smart
object with Lock All or Lock Pixels (its own or a parent group's) took
the distortion and a history step.
Fix: check the pixel and full locks (not position) before recording the
smart filter, as regular filters do, and use the filters' error wording.
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
After Replace Text the find cursor collapsed to the end of the replacement.
A backward search takes the last match ending at or before the cursor, so
when the replacement still matched (e.g. "cat" -> "Cat", case-insensitive)
it found the replaced text again and never reached earlier matches.
Collapse the cursor at the start of the replacement when searching backward
(the end when forward), mirroring how forward search skips it.
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
