#4389, #4390: drag a selected chart's data outline to change its range
A selected chart now outlines its data as Excel does: series names in red,
category labels in purple and the values as one blue block. The values
block has a handle on each corner. Dragging one previews the new block,
and the release points the chart at it as one undo step. Esc cancels.
The new engine command chart.setValuesRange does the move. Each series
keeps its formatting, a wider block adds series and a narrower one drops
them, and names, category labels and bubble sizes that line up with the
block follow it. PivotCharts and series that are not one block are
refused.
#4389: chart data drag review fixes
- Category labels are compared as blocks, so a reference written with a
leading = still counts as the same labels.
- A resize keeps the orientation the series were read in, so a chart
shrunk to one cell per series stays resizable.
- The preview stays while edge scrolling moves the chart off screen.
- A values corner hidden under frozen panes is not a handle.
- The release takes the block from the pointer's last position, even when
it moved in the release frame.
- Esc cancels the drag with Ctrl or Cmd held too.
Each fix has a test that fails without it.
#4389: read a corner drag from the grabbed corner's own cell
A values corner lies on a cell boundary, which belongs to the next row
and column, so a press on the bottom-right corner of A1:A5 read as B6:
releasing where the press was, or dragging straight down the edge of
column A, changed the other axis too. The drag now records the offset
from the press point to the centre of the corner's cell and reads the
pointer through it, so the press point stays in its cell and each whole
cell of movement moves the corner by one.
#4389: snap a dragged chart data edge to the nearest gridline
Reading the pointer from the middle of the grabbed corner's cell let the
grabbed row's height decide when the block moved: with a tall row 5 and
a short row 6, the pointer reached the bottom of row 6 while the block
stayed at A1:A5. Each edge the corner moves now snaps to the gridline
nearest the pointer, so every row and column changes over at its own
middle. A release where the press was still changes nothing, since the
corner lies on its own gridlines, and an edge dropped on its fixed edge
keeps one row or column.
#4389: drop a chart data drag when the chart's data changes under it
A command run between frames (the control channel, a macro) could change
the chart's values block while the button was held, and the release then
put the dragged block over that change. The drag now checks, from the
session, that the chart still plots the block it started from: a move
drops a stale drag, and the release skips it.
XLSX: read, evaluate and write Excel 2010 conditional formats
Excel 2010+ keeps conditional format rules whose formulas refer to
other sheets only in the worksheet's x14:conditionalFormattings
extension, with xm:f formulas and an inline x14:dxf in place of a
dxfId. GridCraft skipped the whole extension with a warning.
Read: x14 cellIs (every operator, one or two formulas), expression,
blanks, errors, duplicate/unique, top 10, above average and time
period rules become ordinary rules, with the inline differential
format (font, fill, border, number format) and priorities that
interleave with the main section's. A data bar's linked extension
(the x14:id of a main-section rule) adds no second rule. Other x14
kinds (data bar settings, x14 icon sets) are skipped with a warning
naming the kind.
Evaluate: cross-sheet references in rule formulas already resolve;
an engine test pins cellIs and expression rules against cells on
another sheet.
Write: a cell-value or formula rule referring to another sheet is
written in the x14 extension ({78C0D931-6437-407d-A8EE-F0AAD7539E65})
with an inline x14:dxf and a unique id; other rules stay in the main
section. The extension shares one extLst with the sparklines (CF
first). Inline number formats get ids styles.xml declares.
XLSX: say which x14 conditional format rules stay in the main section
Colour scale, data bar and icon set rules are always written in the main
section, even when their cfvo formulas refer to other sheets; the parity
docs now say so. (This commit first dropped the
reader's own x14:id linking; after the rebase onto main, main's data bar
linking from #550 is the one path, so only the docs part remains.)
XLSX: correct the x14 conditional format provenance and list skipped kinds
The file that motivated the x14 conditional format work
(watchdog-report.xlsx) was saved by LibreOffice 25.2, not Excel, so
nothing observed in Excel backs the earlier commit's statement that
Excel 2010+ keeps cross-sheet rules only in the extension. MS-XLSX
2.6.27 (CT_CfRule) allows single-sheet references in xm:f but says
nothing about where Excel places such rules; the comments and docs now
say the placement follows the LibreOffice-saved file and Excel's own
is unverified.
MS-XLSX 2.6.27 forbids sheet-range-reference (3-D spans) in xm:f, so a
rule with a 3-D span now stays in the main section.
The parity docs and gaps list name the x14 kinds read (cellIs,
expression, blanks, errors, duplicate/unique, top10, aboveAverage,
timePeriod) and those skipped with a warning and lost on save (data
bars, icon sets, colour scales, containsText, notContainsText,
beginsWith, endsWith).
Styles: differential formats keep their cell protection on save
dxf_xml, which writes both styles.xml dxfs and inline x14:dxf elements, wrote font, number format, fill, alignment and border but never protection, although read_dxf reads it. A rule whose format unlocked or hid cells lost that on save. It now writes protection after border, the order CT_Dxf requires.
XLSX: solid data bars are saved with negatives in the bar colour and no axis, as GridCraft draws them
docs: order parity revision history
docs: record verified x14 conditional format placement
docs: fold the Excel check into the PR's own changelog row instead of a new dated row
docs: the Excel check covered cross-sheet expression rules, not cell-value rules
Fix differential format serialization
Preserve default dxf protection
fix(functions): RATE follows Excel's iteration
RATE used Newton with fallbacks and a bracketing search, so it found roots
Excel doesn't and missed the last iterate Excel returns. It now follows
Excel's solver as observed with synthetic inputs: a secant iteration from
the guess and the guess moved 0.001 toward a larger equation value, at most
200 steps, stopping when a step is under 1e-7 and the equation is within
1e-7 of zero (or as close as its terms allow), with the result more than
1e-7 above -100%. After 200 steps the last iterate is kept if it is within
a looser 1e-4. Subnormal steps flush to zero as in Excel.
RATE(5,0,-100,0) is now -0.999999409145182 like Excel (was #NUM!), and
RATE(4,0,-0.0001,100) is #NUM! like Excel (was 30.62). A guess at or below
-1 is #VALUE!. IRR and XIRR keep their solver.
RATE test: compare with a relative tolerance, not exact equality
ROADMAP: keep the progress log in descending date order
RATE: pin Excel's #NUM! for a fractional nper with a guess near -1
- Import SheetApp in the pivot pane tests.
- Move the keyboard owner check out of keyboard() to stay within the complexity limit.
- Aim the title bar button helper at the button's own node.
Fixes#4473.
Signed-off-by: FredApps <153173854+FredApps@users.noreply.github.com>
* xlsx/csv: cap secondary reader allocations (#2159)
- XML: at most 1M children per element, 16,384 per worksheet row (TooLarge).
- Workbook-wide cap of 16M cell elements.
- Array-formula ranges indexed by row (budgeted), not scanned per cell.
- <cols> writes share a 4M budget.
- At most 4096 sheets; duplicate names deduplicated with a HashSet.
- Pivot: field lists capped, repeated fields used once, labels borrowed.
- CSV: fields past MAX_COLS dropped while reading the record.
- At most 4M shared strings.
- Relationships looked up by id through a HashMap.
* xlsx: close the remaining secondary-reader cap gaps (#2159)
- CSV: characters of a field past column 16,384 are not stored; quotes,
delimiters and line ends in it are still followed.
- Pivot: a field list cut at MAX_FIELDS is reported with a warning.
- Sheet names: the prefix shortens once the (n) suffix passes (999), so
deduplicated names stay within 31 characters.
- Rows dropped for a bad row number still charge their <c> elements
against MAX_CELLS.
- Test Ctx helper gains the fields upstream added.
* #2159: describe MAX_CELLS as the reader's own cap
calc::eval::MAX_CELLS has the same value but limits one array result,
not a workbook.
* #2159: warn on capped pivot cache fields, dedupe pivot names through sets
A cache cut at 16,384 fields is now reported, as the table field lists
are. Axis filters and calculated fields are deduplicated through a set
of lower-cased names instead of a scan per field, which was quadratic in
the 16,384 fields allowed; the first of names equal but for case still
wins.
* links: spawn the Windows opener by absolute path (#2176)
Explorer is now spawned as %SystemRoot%\explorer.exe (falling back to
windir, then C:\Windows; only local drive paths are accepted), so an
explorer.exe planted next to the app or in the CWD can't run. Space, ",
comma and ^ are percent-encoded before Explorer parses its command line.
macOS uses /usr/bin/open; a target starting with - or / is refused on
every platform so it can never become an opener option.
* links: rename opener enum variant (clippy)
`Os::MacOs` -> `Os::Mac` (clippy::enum_variant_names). Also give the
desktop-theme watcher's non-Linux `Link` a non-unit type so
clippy::let_unit_value stops firing on Windows/macOS.
* ui: leave desktop_theme Link alone (covered by #2141)
* #2176: say why the Explorer escape still encodes commas
* Linux: run on native Wayland again by default (revert #165)
winit 0.30.13 reported dropped files only on X11. The egui 0.37 migration
(#2543) pins storytold/winit, whose patch W15 reports file drops through
a per-seat Wayland data device (rust-windowing/winit#4729), so files
dragged from the file manager open on native Wayland too. The XWayland
default is no longer needed.
It also costs a startup crash: winit's X11 backend loads
libxkbcommon-x11, which the Linux packages don't depend on. Where it's
missing (a Fedora toolbox, minimal installs), GridCraft panics before the
window opens: "Library libxkbcommon-x11.so could not be loaded."
Checked on GNOME (Wayland, Fedora 44) with a build of main plus other
proposed changes, run with GRIDCRAFT_WAYLAND=1 (the path this revert
makes the default): an .xlsx and a .csv dragged from Files each open as a
workbook, and the app starts without libxkbcommon-x11 installed.
This reverts commit 6f0e43c21c.
Claude-Session: https://claude.ai/code/session_01Rvu8dWh26U2fsLQMSBQfH8
* test: header AutoFit check measures the text as AutoFit did
table_filter_autofit_works_through_command_and_header_double_click
compared the autofitted column with the engine's text-width estimate,
but in the app AutoFit measures with the installed fonts
(Session::text_width). Where those are narrower than the estimate (a
Mac with Carlito, metric-compatible with Calibri), the estimate
"overlapped" the filter button and the test failed on main. It now
measures with the same function AutoFit used.
* test: the header AutoFit UI check measures with the app's fonts, which the app always installs
The fallback to the engine's estimate never ran: SheetApp::logic installs the measure before
AutoFit. AutoFit on the estimate is covered by the engine's autofit_filter_buttons tests.
* Keyboard: punctuation shortcuts follow the character the layout types (#1384)
On a Croatian or German keyboard Shift+, types `;`, so Excel inserts the
date with Ctrl+Shift+,. GridCraft matched the key's US position and ran
Decrease Font Size. Punctuation keys now go by the character typed
(translated to the US key and Shift that type it), digits keep their
position, and a character egui has no key for (č, ö) still falls back to
the position. ribbon::shortcut reads the same key, so Ctrl+Shift+0
typing `=` unhides columns instead of opening Insert, and Ctrl++ opens it.
* Keyboard: Cmd+Shift+; on a Mac doesn't insert today's date (review)
ribbon::shortcut's Semicolon arm took Shift too, so the US-key mapping turned Cmd+Shift+; (and
Cmd+Shift+. on a Croatian or German layout) into the date. Ctrl+Shift+; stays the time.
* Keyboard: tests for AZERTY and Czech, and for typing while a cell is edited (#1384)
French AZERTY (`;` and `:` unshifted on the US `,` and `.` keys) and Czech (`;` on the US
backtick key, `:` Shift+.) give the date and the time with Ctrl. The layout mapping runs only
in the grid, which reads no shortcuts while a cell is edited; a test pins that: `;` and `:` typed
as Shift+, and Shift+. on a Croatian or German keyboard go into the cell, and the font size stays.
* Fuzzer: undo checks follow the bounded undo history
The undo history drops its oldest steps past 100 entries or 128 MB of snapshots (#1263). The
command-sequence fuzzer still compared "undo everything" with the episode's starting workbook,
so any episode whose snapshots passed the budget failed with "undoing all N steps doesn't give
back the starting workbook" although undo worked (seeds 104, 110 and 163 at 80 steps: a column
inserted next to whole-column borders, or rows deleted after filling whole columns, each made a
100 MB step).
The harness notices when an undoable command changed the workbook without the history growing,
and then compares with the state before the oldest step still in the history.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Fuzzer: undo and redo checks follow the reachable history boundary
Review of #2332. The harness now tracks the oldest state the undo history
can reach instead of a trimmed flag: whenever the history's oldest entry
changes without it emptying in between, the history dropped a step, and
undoing everything stops at the new oldest entry. That boundary is kept
when an edit.undo in the sequence uses up the retained history before the
episode ends, which used to fall back to the episode's start and report a
false undo mismatch.
The redo round trip no longer assumes n undos leave n redo entries: the
redo history is bounded too, so when it dropped the first step undone,
redo is expected to stop at the oldest entry it kept.
---------
Co-authored-by: nihilist-narwhal <240798405+nihilist-narwhal@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* Flush a saved workbook to disk before the rename
write_file renamed the temporary file over the workbook without sync_all, so a power loss right after a save could leave an empty or truncated file under the workbook's name. Sync the file before the rename and, on unix, sync the folder after it (best effort).
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* Tolerate file systems that cannot flush when saving
On Apple targets std's sync_all asks for F_FULLFSYNC, which network mounts refuse with ENOTTY or ENOTSUP. Treat exactly those errors (and ErrorKind::Unsupported) from the temp file's sync as 'cannot flush here' and carry on with the rename, as before the flush was added. All other errors still fail the save.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* Treat raw error 25 as unsupported-flush only on Unix
On Windows 25 is ERROR_SEEK, a real failure that must not be skipped.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* Save: treat Apple's ENOTSUP from F_FULLFSYNC as unsupported flush
On Apple targets ENOTSUP is 45 and std leaves it uncategorized (only
EOPNOTSUPP, 102, and ENOSYS map to ErrorKind::Unsupported), so a volume
refusing F_FULLFSYNC with ENOTSUP failed the save and deleted the
temporary file. The raw code is now matched on Apple targets only;
elsewhere 45 is a different errno and stays a failed save. The check is
split into a platform-parameterised helper so each platform's codes are
tested on any host, plus an Apple-only test on the real helper.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
* Sort: Custom Sort dialog redesign
* Sort: keep the dialog on its worksheet and open when sorting fails
* Improve custom sort dialog and localisation
Localise sort controls and adapt their layout to translations. Track worksheets by identity, restore selections across sheet changes, and allow additional levels to reuse columns.
* Fix sort dialog worksheet and input handling
Open Sort on the requested worksheet, reject invalid ranges, and close any open filter menu so Enter confirms the sort. Add regression tests for each case.
A table whose Total Row was turned off still holds each column's total:
Excel writes totalsRowFunction, totalsRowLabel and a custom
totalsRowFormula without totalsRowCount, and restores them when the row
is turned back on. The reader took them, but the writer wrote them only
while the row was shown, so a save lost a custom formula or an Average
and Excel brought back SUBTOTAL(109,...) instead. They are now written
whether the row is shown or not.
Fixes#4296
A sheet set to First page number 7 has pageSetup firstPageNumber="7"
useFirstPageNumber="1"; neither attribute was read, so a save reset it
to Automatic. PrintSettings now holds it (None for Automatic), the
reader takes the number when useFirstPageNumber is on, and the writer
writes both in the schema's order. A JSON workbook leaves it out when
Automatic, so existing files are written as before.
Fixes#4293