openflo 2.6.2__tar.gz → 2.7.0b2__tar.gz

{openflo-2.6.2 → openflo-2.7.0b2}/CHANGELOG.md

@@ -6,6 +6,923 @@ this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.htm
 
 ## [Unreleased]
 
+This becomes **2.7.0b2**, the first public beta on PyPI (the version in
+pyproject.toml); 2.7.0b1 went only to the public source mirror. It is tagged
+as a pre-release, so `pip install openflo` keeps installing 2.6.2 until
+`--pre` is given, and the update check offers a beta only to someone already
+on a pre-release.
+
+### Changed
+- **No error disappears unseen any more.** An error in a background
+  thread, or in background work with no handler of its own, now opens the
+  same "Something went wrong" window and problem report as any other,
+  instead of only printing or only flashing in the status bar. The 418
+  places that caught every error and did nothing now catch only what they
+  expect, write a "[handled]" line to the log, or say why nothing can
+  matter there; a test keeps new ones out.
+- **Dialogs say which setting is wrong.** Fifteen number settings
+  (Cluster, Cell cycle, Proliferation, Bead tracking, Auto-gate, Batch-norm,
+  Calibration, Embedding, MEM annotation, FMO, Spectral unmixing,
+  Trajectory, Figure layout and the layout editor's text size) refuse text
+  that is not a number in range with a message naming the setting.
+  **Behaviour change:** several used to swap a mistyped value for a default
+  without saying so, or clamp it into range; they now stop instead.
+- **Re-fit per sample finds the valley before a bright population at the
+  top of the range.** It used to say "one population, no valley" and keep
+  the old cut when a tight population sat at the top of the data.
+  **Behaviour change:** such gates now move to the valley.
+- **The layout editor's Layout and Export buttons** no longer show the menu
+  arrow twice.
+- **The embedding warm-up no longer stalls.** Tk clean-up that a background
+  thread's garbage collection triggers is handed to the window's event loop
+  rather than waiting on it.
+- **Group comparisons can come out significant at 3 to 5 samples a
+  group.** Mann-Whitney's smallest possible p at 3 vs 3 is 0.1, so two
+  groups that did not overlap at all read "not significant". Frequencies
+  and Marker expression now default to Welch's t-test (ANOVA for more
+  groups) on transformed values: percentages as log odds, counts as log,
+  intensities on the arcsinh scale (cofactor 150). The Test box keeps the
+  rank test and a t-test on raw values. **Behaviour change:** p-values in
+  those windows, Compare all, and the CLI's `compare_*.csv` (now Welch's t
+  on the logit of each sample's %) differ from earlier versions on purpose.
+- **Paired by donor.** A tick in Frequencies and Marker expression matches
+  samples by their name with the group taken out (D1_Stim with D1_Ctrl):
+  two groups get a paired t-test (Wilcoxon signed-rank with the rank test),
+  more get a repeated-measures ANOVA (Friedman). Differential abundance
+  then models the donor as a term (`count ~ donor + group`, as diffcyt's
+  edgeR model), so a donor's own level is taken out before the groups are
+  compared. A sample with no partner is left out and named.
+- **Saying when a test can't reach 0.05.** Every result carries the
+  smallest p its test could give at those group sizes; the summaries,
+  Compare all and Group comparison say plainly when that is above 0.05.
+- **MESF calibration fits log-log by default.** Tools > Calibration fits
+  log MESF against log MFI, the standard for bead sets spanning decades;
+  Linear is one click away. `fit_mesf_calibration` and `apply_calibration`
+  take `model='loglog'` and keep linear as their default.
+- **Faster `openflo-run` batches, same files.** With `--workers` above 1,
+  each sample's scatter plots and stats CSV are drawn in the worker that
+  clustered it, instead of one at a time in the main process while the
+  workers wait; group UMAPs are drawn in parallel; and `compare` takes
+  `--workers` (default up to 4) and compares samples in parallel, reporting
+  them in workspace order. The files are byte-identical to a one-worker run.
+- **Every dialog behaves the same way.** Each window opened from the editor
+  stays in front of it (no taskbar entry of its own), Escape closes it, and
+  Return runs its main action, which sits beside Cancel or Close at the
+  right of the button row. The layout editor and a popped-out workspace stay
+  separate windows.
+- **Errors in plain words.** An error says what failed and why ("The file
+  may be open in another program") instead of the exception's name; Show
+  details has the technical description and Copy details copies it for a
+  bug report. An unexpected error opens that window once instead of popping
+  the log and console open.
+- **Keyboard shortcuts and About are proper windows**: shortcuts in two
+  columns by menu, and About with the version, licence, how to cite (with a
+  Copy citation button) and links.
+- **Clearer wording.** Side-panel buttons say what they do (Delete gate,
+  Delete all, Copy gates…, Populations…, Undo, Redo, Fit, Apply to all
+  shown), beta tags moved into the tips, no arrows used as words (File >
+  Export FlowJo workspace…), sentence case, American spelling, the methods'
+  published names (PhenoGraph, t-SNE, TriMap), and the Pipeline's fields
+  named in full (Method, Neighbors, Resolution, Max events, Seed).
+- **The title names the session**: "session name — OpenFlo", with * while
+  there are unsaved changes.
+- **Menus regrouped, with their shortcuts.** Tools is in four groups
+  (compensation and scales, figures, utilities, templates); exports moved
+  to File, and Methods & provenance sits beside History / audit in Analyze.
+  Menus show Ctrl+Z, Ctrl+Y, Ctrl+F, the zoom keys and F5, and on macOS
+  the same keys answer to Command. F1 opens the documentation, and the main
+  window's File menu ends in Quit OpenFlo (Ctrl+Q).
+- **No pop-up after an export.** The status bar says what was saved, with
+  an Open folder button, instead of a box to click away.
+- **Theme: windows and plot background, chosen separately.** View > Theme
+  is Light or Dark windows, and View > Plot background is White or Black;
+  figures in analysis windows follow the plot background, so "Dark figures
+  in pop-ups" is gone. Sessions keep their theme.
+- **Readable fonts and colours.** Windows use named fonts that follow the
+  window-width scaling and are never below 9 pt. Error, warning and success
+  text, the drag highlight and the log pane take their colours from the
+  theme, each readable at 4.5:1 or more; the log is light in the light
+  theme.
+- **Cluster dialog**: only the picked method's settings show (FlowSOM's
+  grid and metaclusters, Leiden's resolution), and markers are listed as
+  "marker (channel)", as the axis menus show them.
+- **Layout editor**: Ctrl+Z and Ctrl+Y undo and redo, Ctrl+S saves the
+  session (which carries the layout), and a change marks the session
+  unsaved, so closing asks to save it. The empty-page hint names the real
+  buttons and the status line wraps instead of running off the window.
+- **Tools > Quick figure grid** is the old Figure layout, one name in the
+  menu and the window, so it is not mistaken for the layout editor.
+
+### Added
+- **Results record the OpenFlo version that made them (session schema
+  v3).** A gate placed by Re-place on each sample or by Auto-gate, and each
+  column written by clustering, an embedding, cell cycle, calibration or
+  pseudotime, carries a stamp: the step, the OpenFlo version and what it ran
+  on. The `.flowsession` gains a top-level `app_version`, and each sample
+  entry an `app_version` and `result_stamps` (gate stamps travel on the gate
+  as `stamp`). Saving writes an older result's stamp back unchanged, so a
+  later fix can find the results made before it. Each audit-trail entry
+  also records the version (`version`). Version-2 sessions upgrade on open,
+  their results' versions unknown; a session saved now does not open in an
+  older OpenFlo.
+- **Fixes reach saved sessions, once and only when you say so.** When a
+  session opens, OpenFlo lists the saved results made by a version with a
+  bug since fixed: each sample's gate or column, the version that made it,
+  and its numbers before and after (events and % of parent for a gate).
+  Tick the ones to correct and Apply (one undo step); nothing changes
+  otherwise. A corrected result is recomputed with this version, its stamp
+  moves to it, and its original stays in the session (a gate's
+  `original`, a column's values under `<column> [original <version>]`,
+  out of plots and statistics). Fixes not taken are recorded and not
+  offered again; Tools > Review fixes for this session… lists them all.
+  Gates and other results built on a corrected one are offered as a
+  follow-up re-run, never started by themselves; for clusters, embeddings,
+  trajectory, cell cycle and calibration, Open… opens that tool on the
+  sample. The same list names the files exported from a corrected sample
+  before the fix (when made, by which version, whether still there), with
+  Show in folder; they are not re-exported. Plot images, Gating-ML,
+  FlowJo .wsp, population FCS folders and the Statistics, Frequencies,
+  Kinetics, workspace-comparison and differential-abundance exports are
+  recorded for it too. A changed default is
+  flagged, not applied. The first fix listed is the re-fit valley fix
+  (#78), for gates re-placed per sample before 2.7.0b2. Exports name the
+  fixes applied. The session gains `fixes_applied` and `fixes_declined`,
+  and each sample entry `result_originals`.
+- **`openflo-run check-sessions <folder>`** lists, for every `.flowsession`
+  under a folder, the fixes that would apply to it. It only reads.
+- **A gate moved by hand is yours.** Dragging, typing coordinates, the
+  arrow keys or a slider takes the step's stamp off the gate (its stamp
+  records the geometry it was placed with), and a pasted or copied gate
+  carries none, so a later fix never overwrites a hand edit.
+- **A reopened session shows its gate counts and statistics at once.** Each
+  sample entry saves the counts its gates had and the Statistics window's
+  last few tables (`saved_stats`, numbers only), and a reopen uses them
+  instead of computing them again, but only when the sample's data digest,
+  the gates (and their ancestors), the statistics chosen, the Preferences
+  that change a formula and the OpenFlo version are all the same; anything
+  else is computed again as before. Measured on 6 samples of 500,000
+  events with 10 gates each: the counts were all shown 0.8 s after the last
+  sample loaded, now as each one loads; the Statistics window was ready in
+  1.1 s, now 0.25 s. Plots are not saved.
+- **Normalize batches dialog.** Tools > Normalize batches (CytoNorm)
+  replaces the one-click batch-norm: batches by trial / day or by any FCS
+  keyword ($DATE by default), the mapping learnt from a reference sample
+  in every batch (named with a token such as "ref"; CytoNorm's standard
+  form) or from all samples (CytoNorm 2.0), 1 to 40 metaclusters, applied
+  to the loaded samples and/or written as normalized FCS copies. It shows
+  each batch before and after on a channel, the per-batch medians and a
+  methods sentence, which also goes into the audit trail. Compensation
+  controls and set-aside samples stay out of the fit. Restore original
+  values (in the dialog, or Tools > Undo batch normalization) puts the
+  loaded samples back on their pre-normalization values without reloading.
+- **Excel workbooks and European CSV.** Statistics has Save as Excel: the
+  whole table on one sheet, then one sheet per statistic (populations
+  down, samples across), and an About sheet with the OpenFlo version.
+  Frequencies can save its tidy table as .xlsx, and `openflo-run --xlsx`
+  collects the run's CSV tables into one workbook. Preferences > Export >
+  CSV tables can follow this computer's list separator and decimal mark
+  (semicolon and decimal comma across most of Europe), so Excel opens the
+  CSV in columns; copied tables follow the same setting. The default CSV
+  is unchanged.
+- **Set a sample aside.** Right-click a sample > Set aside from
+  comparisons, with a reason (failed QC, clogged, too few events). It stays
+  in the session and in the Statistics table (with a Set aside column),
+  is greyed with its reason in the tree, is left out of every group
+  comparison (Frequencies, Expression, Group stats, normalization) and is
+  listed with its reason in the methods text. A free-text note per sample
+  too. Both are saved in the session (additive keys an older OpenFlo
+  ignores).
+- **Colour-blind-safe colours and keyboard gating.** Samples and groups
+  now take the Okabe-Ito colours, which stay distinct under the common
+  colour-vision deficiencies (tab10's red and green do not); View > Group
+  colors switches back to the classic palette, keeping colours you picked
+  by hand. The pipeline's `--palette auto` follows. The plot takes keyboard
+  focus (click it, or F6 from the gate list): the arrow keys move the gate
+  selected in the list, Shift for larger steps, one undo step per run of
+  presses.
+- **Setup without Python.** `setup.bat` and `setup.sh` no longer stop when
+  no Python 3.11+ is installed (an older Python is now refused up front
+  instead of failing in pip). They use a `uv` on PATH, or download a pinned
+  uv release (Windows x86_64/ARM64, macOS Intel/Apple silicon, Linux
+  x86_64/aarch64) into `.tools/uv`, check its SHA-256 against the hash in
+  the script and refuse to continue on a mismatch, then let uv install
+  Python 3.12 and create `.venv`. An interactive Windows setup also adds
+  Desktop and Start-menu shortcuts. The scripts are not code-signed, so
+  SmartScreen or Gatekeeper may still warn.
+- **Batch normalization helpers (library).** `openflo.batch_normalize`
+  wraps the existing CytoNorm model for acquisition-day batches: group
+  samples by any FCS keyword (`$DATE` by default in a dialog), fit on a
+  reference sample run in every batch ('controls') or on all samples
+  ('goal'), apply to new frames without touching the loaded ones, a
+  per-channel, per-batch median and robust-SD before/after table with a
+  methods-report sentence, and normalized FCS copies (compensated linear
+  values, source keywords kept, `OPENFLO_NORMALIZATION` / `OPENFLO_NORM_*`
+  provenance keywords added) written to a folder. Original files are never
+  modified.
+- **Combination gates.** Edit > Combination gates… (also on a gate's
+  right-click menu) takes 2 to 6 gates and makes every combination of
+  them, all positive to all negative, as boolean populations under the
+  gates' common parent, on every sample that has those gates; a negative
+  means in the parent and outside the gate, and an event with no value is
+  in no combination. A polyfunctionality window shows SPICE-style bars
+  per combination (group means with a dot per sample), the +/- grid and
+  pies of how many gates each positive event is in, and exports the
+  frequency table (NA where a frequency cannot be computed). The
+  combinations are ordinary gates, so they reach statistics and the
+  FlowJo export.
+- **Typed gate coordinates.** A gate's right-click > Edit coordinates…
+  types a threshold, interval, rectangle, quadrant divider, polygon's
+  vertices or ellipse's centre and shape in the axis units (intensity;
+  1.5K reads as 1,500, a blank bound is open), optionally on the same gate
+  in every other sample. One undo step.
+- **Gates that re-fit per sample.** A threshold or interval gate, or a
+  singlet or Gaussian-mixture gate from Auto-gate, can be re-placed on
+  each sample's own data (right-click > Re-place on each sample…): a cut
+  moves to the density valley nearest where it was drawn, and a proposal
+  is fitted again by its own rule. A grid shows every sample's old and
+  new placement before you accept; a sample with no landmark keeps its
+  gate and says why.
+- **Compare FMO thresholds.** Edit > FMO gating has a Compare
+  thresholds button: every stained sample's % positive on each mapped
+  channel with the cutoff at each FMO percentile you list (99 and 99.5 by
+  default, i.e. 1 % and 0.5 % of FMO events above the gate), inside the
+  selected parent gate, without adding a gate. Copy or export the table.
+- **FMO on histograms.** Once FMO gating has mapped a channel to its FMO
+  control, a histogram of that channel draws the FMO in grey behind the
+  sample, and the legend gives each one's % at or above the channel's
+  threshold gate. View > FMO on histograms turns it off. The mapping is
+  saved with the session.
+- **Gating strategy page.** The layout editor's ＋ Strategy (also on the
+  page menu) adds a page with one plot per gating step, each the parent
+  population on the gate's axes with its gates drawn; with a gate selected
+  in the main window, only the steps down to it. Batch export draws it for
+  every sample.
+- **PowerPoint export.** The layout editor exports a page, or every
+  sample's pages in a batch, as a .pptx deck: one slide per page, the page
+  drawn at the export resolution and filling a slide of the page's size.
+- **Compare with saved session.** File > Compare with saved session counts
+  every population with this session's gates and with a saved session's
+  gates, on the events loaded now, and highlights each count that moved
+  and each gate that changed.
+- **Gate list.** File > Export gate list (CSV) writes every sample's gates
+  with their population path, kind, channels and coordinates, both as
+  stored and in linear units.
+- **FlowJo export checks Time gates.** A Time gate whose upper edge ends
+  within the first `$TIMESTEP` fraction of the run (where an edge already
+  converted to seconds lands) now warns, with the events it cuts, since
+  OpenFlo keeps Time in ticks and converts it to seconds for FlowJo.
+- **Derived channels.** Tools > Derived channels makes a channel computed
+  per event from others, such as the ratio of two detectors for a ratio
+  dye, FRET, or one marker over another: `[FL1-A] / [FL2-A]`,
+  `log10([FL1-A] / [FL2-A])`. Channels go in square brackets (a marker name
+  works too); + - * / ** and log, log10, log2, sqrt, exp, abs, asinh, min
+  and max can be used, and nothing else is run. It is computed on
+  compensated linear values, shows as `D:<name>` to plot, gate and
+  summarise, is saved with its formula (in the sample entry, so older
+  sessions open as before) and recomputed when a session reopens from the
+  FCS files. An event with no finite result (0 / 0) is NaN, not 0.
+- **Kinetics.** Analyze > Kinetics follows a channel over acquisition time
+  (calcium flux, dye uptake): one binned median (or mean) trace per sample,
+  overlaid, and a table of each sample's baseline, peak after the
+  stimulus, time to peak, area over the baseline and peak / baseline, with
+  CSV and figure export. Time is in seconds when the file gives
+  `$TIMESTEP`, else in ticks, and the window says which. What can't be
+  computed (no baseline events, a baseline at 0 for the fold) reads NA.
+- **Two windows can't overwrite one session silently.** Opening or saving a
+  session by name writes a small lock file beside it
+  (`name.flowsession.lock`) naming the computer and user. A second person,
+  or a window of another OpenFlo, is told the session is open and picks
+  Open read-only (Save then asks for a new name), Open anyway or Cancel. A
+  lock left by a crash on this computer is taken over without asking.
+  Before Save, or Save and close, writes over the file this window opened,
+  it checks the file is still the one it read or last saved; if someone
+  saved in between (or the file was deleted), it asks: Save as a copy,
+  Overwrite or Cancel. Synced folders copy lock files late, so this check
+  is what protects the data.
+- **Every test proven to pass and to fail.** `scripts/prove_tests.py` runs
+  each test on the real code, then breaks the code it guards one change at
+  a time and records the break that made it fail. `docs/test-assurance/`
+  holds the result for all 3,420 tests (3,406 proven both ways) and the
+  method. Tests that needed a developer's own sample files now generate
+  their data, and tests that could never fail were strengthened.
+- Hover help on the auto-gate method choices, the embedding method boxes and
+  the axis-settings parameter box.
+- **Offline user guide.** Help > User guide (F1) opens a guide and
+  tutorial that ship with OpenFlo, so it works without a connection: ten
+  topics from a first analysis to troubleshooting, with search, and the
+  tutorial can load the example dataset. Help > Online documentation still
+  opens the web page. A test checks that every menu path the guide names
+  exists.
+- **Dean-Jett-Fox cell cycle.** Analyze > Cell cycle fits G1 and G2/M peaks
+  and a polynomial S phase (Fox 1980) by default and reports their areas,
+  so S-phase cells under the peaks count as S (synthetic truth 56/30/14
+  read 55.8/30.5/13.6, against 53.4/29.3/17.4 by boundaries). The window
+  draws the fit and lists the boundaries result beside it; Boundaries stays
+  available, and `cell_cycle(method=...)` defaults to it.
+- **Proliferation.** Analyze > Proliferation fits a dye-dilution histogram
+  as generations one halving apart and reports cells per generation,
+  precursor frequencies, percent divided and the division, proliferation
+  and expansion indices (Roederer 2011). The undivided peak is found, or
+  fixed at an unstimulated control's median. Too few cells read NA.
+- **Spreading matrix for compensation.** Tools > Spreading matrix measures
+  the spillover spread of conventional compensation from compensated
+  single-stain controls, matched to their channels automatically; a
+  channel without a control reads NA. CSV and figure export.
+- **Several autofluorescence signatures.** Spectral unmix can find up to
+  five autofluorescence spectra in the unstained control; each cell is
+  unmixed with the one that fits it best, and a U:AF index channel says
+  which. Spectral QC adds Unmixed pairs, an NxN plot of every pair of
+  unmixed fluors.
+- **Bead tracking.** Tools > Bead tracking keeps a Levey-Jennings log of QC
+  bead runs (a CSV) and judges each run by the Westgard rules (1-3s, 2-2s,
+  R-4s, 4-1s and 10x reject; 1-2s warns).
+- **Index-sort data.** Index-sort locations in the FCS keywords
+  (INDEX SORTING LOCATIONS, $INDEX and similar) are read, and the plate view
+  shows each sorted cell in its well, coloured by a channel.
+- **Group samples by keyword or table.** Edit > Group samples puts the
+  samples into groups by an FCS keyword the files carry (tube, condition,
+  plate, ...) or by a sample-to-condition table (CSV or tab separated: the
+  sample or its file name, then its group), with a preview before Apply.
+  The groups are the ones the sample tree and the group statistics use.
+- **Plate view.** Analyze > Plate view shows a 96-well plate heatmap of the
+  events, or a gate's count or percent of its parent, per well. The well
+  comes from the well keyword or the sample's name (Plate1_B07); a sample
+  without the gate reads NA, and samples without a well are listed.
+- **Export populations, all samples.** File > Export populations, all
+  samples (FCS) writes every sample's gated populations, a folder per
+  sample.
+- **Concatenate samples** from File: joins several samples with the same
+  channels and scales into one new sample, each event tagged with its
+  source in a `__sample__` column.
+- **Downsampling the data can be undone.** Setting the downsample from
+  Display + data back to Display only or Off restores every event, except
+  in a sample whose data changed after the trim, which is named.
+- **More statistics.** Analyze > Statistics adds %Grandparent, SD, robust
+  SD (half of P84.13 to P15.87), MAD, any percentiles you type (P5, P95,
+  ...) and the MFI ratio: a population's median over the same population's
+  median in a reference sample you pick, such as the unstained control.
+  The column ticks sit on three rows so they fit a laptop screen, and Copy
+  table puts the table on the clipboard, tab separated, for Excel or Prism.
+- **Editable figure text.** PDF and SVG exports keep their labels as text
+  (TrueType in PDF, `<text>` in SVG), so Illustrator and Inkscape can edit
+  them; Preferences can switch it off. Preferences also set the export
+  resolution (150 to 1200 dpi, 300 by default) and the figure font, from the
+  fonts installed. File > Copy plot to clipboard, and right-click on the
+  figure in analysis windows, copy a figure as an image.
+- **Saturation report.** Analyze > Saturation report lists, per sample,
+  the percent of events at the top of each channel's range (`$PnR` - 1, on
+  the raw detector values) and across any channel, flags channels at 1 %
+  or more, and exports or copies the table. A channel whose range the file
+  does not give reads NA.
+- **Gating-ML 2.0 files.** File > Export Gating-ML 2.0 gates writes the
+  active sample's gate tree in the ISAC standard other analysis tools read,
+  with each channel's scale (logicle, hyperlog, asinh as `fasinh`, log as
+  `flog`) and the sample's compensation matrix, so polygons keep their
+  shape. Templates > From file… and `openflo-run --gates-from` read them;
+  a polygon stays straight on the axes the file names. Gates Gating-ML
+  cannot hold, or OpenFlo cannot read (quadrant gates, ratio dimensions),
+  are left out with what is under them, and named.
+- **FlowJo groups come across.** Opening a `.wsp` puts each sample in an
+  OpenFlo group named after its FlowJo group (not 'All Samples' or
+  'Compensation'), and the `.wsp` export writes OpenFlo's groups as FlowJo
+  groups. A workspace whose FCS files moved finds them in the subfolders
+  of its own folder or the FCS folder; two files of one name are told
+  apart by their folder, never guessed.
+- **`openflo-run` reproducibility.** Each run writes `run_manifest.json`
+  (command, settings, software versions, gates, and inputs and outputs
+  with SHA-256). `--gates-from` gates every sample on one population
+  (`--gate-population`) of a GUI session, template, `.wsp` or Gating-ML
+  file. `--gate-units data` takes `--gates` in intensities instead of
+  logicle units. The comparison CSV gains `p` (Mann-Whitney U) and `p_adj`
+  (Benjamini-Hochberg), as the GUI's differential abundance has.
+- **Project archives.** File > Export project archive saves one .zip with
+  the session, its saved data and every sample's FCS file, to send or move
+  as one file. Open session opens it directly: it unpacks beside the zip
+  and relinks the samples there. Samples whose FCS file was not found are
+  listed when exporting.
+- **Save session as.** Save (Ctrl+S) now writes back to the session you
+  opened or last saved, and asks for a name only the first time; Save
+  session as picks a new one. The window title shows the session's name.
+- **Exploded view of a zoom or gate in layouts.** A layout plot whose zoom
+  hides part of its data shows the whole plot with the zoomed region
+  outlined, beside a magnified copy joined to it by connector lines, on the
+  page and in Export page and Export batch. Edit… sets it per plot (Auto,
+  Always when zoomed, Off), chooses the copy's side, and can magnify one of
+  the plot's gates (its events, outline and %) instead of the zoom.
+
+### Changed
+- **Nothing written inside the installed package.** PhenoGraph's scratch
+  files now go to a temporary folder per run that is removed afterwards, even
+  when the run fails (they used to collect in a `.phenograph_cache` folder
+  inside the package, which is read-only on a system Python). Saved gate
+  templates now go to `~/.openflo/templates`; templates saved by an older
+  OpenFlo inside the package still show in the Templates menu.
+- **The editor window appears sooner.** It paints before matplotlib loads
+  and the plot area fills in right after, so the window shows in about
+  0.3 s instead of 0.7 s on a warm start (more on a cold disk).
+- **`openflo-run --help` answers at once.** Help and argument errors no
+  longer load the scientific stack (0.07 s instead of 0.8 s); the console
+  script's entry point is now `openflo.cli_args:main`.
+- **The window keeps answering while OpenFlo works.** Dragging a gate
+  redraws only its outline, and its counts are made in the background
+  after you let go (they used to be redone on every mouse move with gates
+  applied). Zooming or changing axes no longer rebuilds the Samples & Gates
+  tree. Trajectory, Marker expression, FMO gating, cell cycle, population
+  export, Sample QC and compensation optimisation run in the background
+  with a progress bar and Cancel; layout batch export shows how far it got
+  and stops on Cancel. Results are unchanged.
+- **Gate rows in the Samples & Gates tree show the population.** Each row
+  reads the gate's name, its event count and its frequency of the parent
+  ('Cells  2,796 · 69.9%') instead of its bounds, which were cut off at the
+  panel's width. An unnamed gate is named by its channels ('FSC-A x SSC-A')
+  or its cut ('FL1-A >= 1K'). The tree widens to its longest row and
+  scrolls sideways. A population that cannot be computed reads NA.
+- **Colouring by cluster gives each cluster its own colour**, the colour of
+  its population in the tree, with a legend naming them. It used a
+  continuous scale, so neighbouring clusters looked alike.
+- **Groups are named after the sample's own folder** when no 'Day N' folder
+  is above it: `synth/pbmc/` and `synth/compensation/` are two groups, no
+  longer both "synth", and the example data is no longer ".openflo". A
+  generic 'Blank Experiment with…' folder still defers to the one above it.
+- **Samples are listed by name** once loading is done, not in the order
+  they finished loading.
+- **Gates imported from a FlowJo workspace come in ticked**, like gates
+  drawn in OpenFlo. They came in unticked.
+- **The Add FCS and Load CSV dialogs open in the folder last picked from,**
+  else the current folder, else the home folder; never the installed
+  package's own folder.
+- **Plots are drawn the way flow papers draw them.** Scatter (FSC, SSC)
+  and Time open on a linear axis, so the cells no longer sit squeezed into
+  one corner of a log plot. Fluorescence opens on logicle, so events at or
+  below zero keep their own positions near 0 instead of piling up on the
+  axis. Plots open in pseudocolor, coloured with Turbo, a rainbow that is
+  close to perceptually uniform (View → Density colors switches back to the
+  classic rainbow or to Viridis). A channel whose scale a session already
+  saved keeps it.
+- **Logicle width and arcsinh cofactor can be set per axis.** The axis
+  dialog (⚙ beside X and Y) offers Logicle, Arcsinh, Log and Linear, with
+  logicle's width W and arcsinh's cofactor. These change only how the axis
+  is drawn: gates select the same events, and the data is not
+  re-transformed. The setting is saved with the session, its tabs and
+  layout items.
+- **Gates show their name and frequency on the plot.** Each gate is drawn
+  with a solid outline and labelled with its name (when it has one) and its
+  percent of the parent, one decimal. Quadrants are labelled in their outer
+  corners, and a population that cannot be computed reads NA. View → Gate
+  labels on plots turns the labels off.
+- **One notation for axis ticks and percentages.** Every decade on an axis
+  reads as a power of ten (10², 10³, not "100" next to "10³"). Linear axes
+  read 50K and 100K. Frequencies in the Statistics table and the status bar
+  have one decimal (93.0, not 93 or 92.96). A gate whose parent is empty
+  reads NA instead of 0.00%.
+- **No window hides its buttons.** Statistics, Methods & provenance,
+  Absolute counts, Preferences, FCS inspector, Expression, Frequencies,
+  Calibration and Voltage optimization opened at fixed sizes too small for
+  their content, and the button row (or Update, or Detect peaks) was simply
+  not shown. Every window now opens big enough for all its controls (on a
+  scaled display too), centred on the main window and on the screen; a
+  window made smaller shrinks its table or plot, not its buttons.
+  Preferences scrolls when the screen is too short and lists themes by
+  name; the layout editor's toolbar wraps onto two rows when narrow.
+- **Statistics tells a parent from its child.** The Population column
+  shows each gate's own name, indented under its parent, in a column as
+  wide as the names; the exported CSV keeps the full path. Statistic
+  columns are as wide as their headings.
+- **Sharp text on scaled Windows displays.** OpenFlo declares itself
+  DPI aware, so at 125 % or 150 % Windows no longer stretches a blurry
+  96-DPI image of every window.
+- **One answer to "is this engine available?".** Help → Environment,
+  Compare embeddings, the Cluster dialog and the Pipeline all test an
+  engine by importing it (in a separate process, once per session), so a
+  package installed without its own dependencies is shown as "fails to
+  load", with the reason, everywhere, instead of being offered and then
+  producing nothing.
+- **No stray untitled box over a window.** It was a hover tip left behind
+  when its window closed, or was busy, under the pointer. Tips now close
+  with their window and are not shown once the pointer has left.
+
+### Added
+- **Compensation can be applied to the whole acquisition batch.** Apply
+  reached only the highlighted sample, so a set of files took one Apply
+  each, and its tooltip said "every loaded sample". The Compensation window
+  now has "Apply to: this sample / this batch". The batch is the samples
+  acquired like the highlighted one, read from each file's keywords rather
+  than its folder: the same instrument (`$CYTSN`, else `$CYT`), acquisition
+  date (`$DATE`) and detector voltages (`$PnV`) on the matrix's channels,
+  and the same stored spillover matrix when both files carry one. The
+  window lists the batch, and hovering it says why each other sample was
+  left out (other voltages, another date, no keywords). A file without
+  these keywords is never guessed into a batch.
+- **Tabs in the Samples & Gates tray (beta).** A tab is a saved view: its
+  samples and gates, their on/off states, and the plot's axes, type,
+  colouring, axis types and zoom. Selecting a tab shows and plots only its
+  members and restores its plot; "All" always holds every sample. ＋ makes a
+  tab from the selected rows (or empty); right-click a tab to rename,
+  duplicate or delete it. In the tree's right-click menu, "Copy to tab" adds
+  the selected samples or gates to a tab (a gate brings its ancestors and
+  subtree), and "Apply to other tabs" (all, or chosen) pushes their on/off
+  state to other tabs. "Within a tab (choose samples)…" applies it to only
+  the samples you find and tick in one other tab, matching gates by
+  population name. Samples and gates can also be dragged onto a tab on the
+  strip: an arranged tab (samples from several groups, or its own order)
+  holds them in a Presort zone at the top of its tree, which grows as needed
+  and vanishes when empty; drag a row out of it to place it. Any other tab
+  takes them straight to the top of its stack. Items dragged back from the
+  Pipeline onto the tree or a tab join that tab. The main plot keeps your
+  zoom through every redraw of the same view (a gate edit, a toggle, a tab
+  switch) until Reset, and each tab keeps its own. Gates are shared: an edit shows in every tab holding
+  that sample. The tabs pop out with the tray. The layout editor's "last
+  displayed" view also reads the views saved in tabs.
+- **Sessions save the tabs and the layout (schema v2).** `.flowsession`
+  gains `view_tabs` and `layout`; version-1 sessions upgrade on open (one
+  All tab, no layout). A session saved now does not open in an older
+  OpenFlo.
+- **Layout editor (beta).** The side panel's Layout button (was "to be
+  added") opens a FlowJo-style editor of pages: plots of gated populations,
+  stats tables (Count, %Parent, %Total and per-channel Median / Mean /
+  GeoMean / CV / rCV), violin / box / bar charts across samples, and text
+  boxes. Drag to move, drag the corner to resize, double-click to edit; an
+  optional grid, and soft snapping to grid lines and to other items' edges,
+  corners and centres, with guides. Samples and gates dragged from the
+  Samples & Gates tree or the Pipeline (docked or popped out) land as plots,
+  shown as that population was last displayed (Preferences can make it the
+  current graph instead), or join the plot, table or chart they are dropped
+  on, so several samples and populations overlay in one plot. A plot keeps
+  the axis types and zoom it was taken with; when much of its data falls
+  outside, a Zoom to fit badge offers (never applies) a refit. Cluster
+  results (PhenoGraph, Leiden, FlowSOM) colour a plot by cluster and their
+  populations can be plotted and tabulated. A View zoom (Fit, 50-200%,
+  custom) sets the on-screen resolution; Export DPI defaults to 300 (150,
+  600, 1200 or custom). Pages can be added, renamed, duplicated, items copied
+  or moved between them, and pages imported from or sent to another
+  project's layout file. The layout re-applies to every sample, matching
+  populations by name; Batch export… writes every page per sample. The
+  mouse wheel scrolls the page (Ctrl+wheel zooms it, Shift+wheel scrolls
+  sideways); over the selected plot it zooms that plot around the cursor.
+  Holding the middle
+  button drags its view, and right-click has Center zoom here, Zoom in,
+  Zoom out and Zoom to fit; the view is kept with the plot. A cross-page
+  host in the left margin keeps copies of items dragged onto it, to drag
+  onto any page or place on every page; each is listed by number, kind,
+  populations, axes and the page it came from, and can be renamed. Drag
+  the host's edge to widen it; pointing at an entry outlines the items on
+  the page that are still true duplicates of it (same populations, axes,
+  zoom, colours...) and marks (◆) the tabs of other pages that hold one.
+  Every item has a lasting id (P1, S2, C3, T4), shown in its corner on
+  screen and in the host, so look-alike plots can be told apart. The toolbar reads left to
+  right: insert, edit, the sample shown, Export (its dpi on the button, 300
+  by default) and the Layout menu; the row below holds the page tabs, page
+  size, grid, snap and the on-screen zoom. A plot dragged
+  onto a stats table or chart supplies its samples (and its populations, if
+  it has none). Each layer's colour can be set per item; a default colour
+  list (layer 1 red, 2 blue, ...) fills in, and an override toggle (off by
+  default) makes every layer use that list. A population that cannot be
+  computed reads NA in tables, gets no bar (NA instead) and no violin, and
+  its plot says so; a measured empty population still reads 0. Layouts
+  save to `.oflayout` files (version 2; version-1 files open). Also under
+  Tools → Layout editor (beta)….
+
+### Changed
+- **Saving and reopening sessions is much faster.** A session keeps each
+  analysed sample's events beside it. They were CSV text, and formatting
+  every value as text was most of a save: about 10 s per 300,000-event,
+  24-channel sample, holding the windows for the 5-minute autosave. They
+  are now binary (`<session>_data/<sample>.npz`, every value and type kept
+  exactly): four such samples save in 0.5 s instead of 44 s, reopen in
+  0.6 s instead of about 6 s, and take 58 MB each instead of 136 MB.
+  Sessions with CSV sidecars still open, and are converted on their next
+  save. A column the binary format can't hold (a nullable or categorical
+  one) keeps that sample on CSV.
+- **The 5-minute autosave no longer freezes the windows.** It now writes on
+  a background thread, as the close does; a close or Save started while it
+  is writing waits for it.
+- **The first save after opening a session no longer rewrites every
+  sample.** The sidecars the session was opened from are remembered, so the
+  autosave and the close copy an unchanged one instead of encoding it again.
+- **Faster analysis with identical results.** Each of these gives the same
+  numbers as before, bit for bit, checked against the previous code:
+  - FlowSOM training (and CytoNorm, which uses it) runs compiled with numba,
+    already installed with UMAP: about 14 s to 1.2 s per sample at the
+    defaults.
+  - The Statistics table splits large tables across CPU cores and computes
+    only the statistics asked for: 31 s to 10 s for 2 million events, 20
+    gates and 16 channels on 4 cores, faster with more.
+  - Leiden and Louvain neighbour searches run on several cores: 7.8 s to
+    2.6 s for 50,000 events on 4 cores. Ties between equally near events
+    resolve exactly as before.
+  - Exporting a population as FCS converts the events in one step: 9.4 s to
+    1.5 s for 2 million events and 16 channels, with about 1 GB less memory,
+    and a byte-identical file.
+  - Integer FCS files that store fewer bits than their words hold (10-bit
+    data in 16 bits, for example) load about 70x faster: 8.8 s to 0.12 s for
+    2 million events and 20 channels.
+
+### Removed
+- **Out-of-date developer files.** DESIGN_NOTES.md (last checked against the
+  code in July, before most of the current interface) and five one-off
+  scripts nothing used: four speed benchmarks and mutate_controls.py.
+
+### Fixed
+- **The public mirror's CI no longer fails its privacy gate.** That job
+  and the release build's gate step ran scripts/publish_public.py, which is
+  kept out of the mirror because it holds the private token list. Both
+  now skip on the mirror; the gate still runs here before every snapshot
+  and every PyPI upload.
+- **Compare embeddings shows Run and Cancel on a 150 % display.** The
+  window took a fixed 420x360 and packed its buttons last, so larger text
+  squeezed the button row to half height on Windows. It now takes the size
+  its content asks for.
+- **New windows stay on a busy machine's screen.** On a loaded Windows
+  PC a window could be fitted on opening and then put back where Windows
+  first placed it, past the bottom of a laptop screen (seen with Plate
+  view at 1366x768). The editor checks the place again just after.
+- **Help > Report a problem reaches the issue tracker.** It, and Help's
+  documentation link, opened the private development repository, a 404 for
+  everyone else. They open openflo-public now, and a test fails on any link
+  to the private repository in what ships.
+- **FCS files with common writer quirks open.** A file whose $ENDDATA is
+  one byte off, whose HEADER and TEXT disagree on where the data starts, or
+  that chains a second data set with $NEXTDATA was refused and reported as
+  "cut short" or "damaged". Each now loads (the first data set, for a
+  chained file) with a load warning naming the quirk; the compensation
+  readers and the FCS inspector read them too. A file that is really
+  damaged still fails, with a message that names the quirk when one is the
+  cause.
+- **Clustering and the FlowSOM tree repeat with their seed.** igraph has no
+  number seed: `layout_fruchterman_reingold(seed=)` takes starting
+  positions, so the tree layout was drawn differently every time. It now
+  runs on a generator seeded from `seed`. Louvain left igraph on a fresh,
+  unseeded generator after each run, so `random.seed()` stopped controlling
+  igraph for the rest of the session; it now hands back the default.
+  PhenoGraph's seeded Leiden backend stays the default, as before.
+- **Leiden noise reads as noise.** Its unassigned bucket imported as
+  "leiden -1" in a palette colour and the Populations menu left Leiden out;
+  it is grey and named "Unassigned (noise)", and the menu lists it (other
+  Leiden populations keep their names, such as "leiden 3").
+- **`python -m openflo.inspect_fcs` lists the channel keywords** again
+  (flowio 1.x returns them lower-case, without the `$`).
+- **The update check never offers a beta to someone on a final release**,
+  and without the `packaging` module a beta no longer sorts above its own
+  final release.
+- **The README's licence note** said OpenFlo does not import igraph and
+  leidenalg (both GPL) directly; it does, and the note says so.
+- **The Cluster dialog starts FlowSOM at 20 metaclusters, and says why.**
+  Asked for exactly as many metaclusters as there are populations, FlowSOM
+  merged a rare (1.5%) population into a neighbour (F1 0.22; 0.99 with 12).
+  The dialog now starts at 20 (was 10), notes "set above the populations
+  you expect" beside the field, and its hover tip explains the reason and
+  how to join extra metaclusters with an OR gate. **Behaviour change:**
+  FlowSOM runs from the dialog at its default now make 20 metaclusters, not
+  10. The CLI and the Python API keep 10.
+- **A 30-channel recovery test for every clustering method.** Eight planted
+  populations in a lineage tree, one of them rare (1.5%): PhenoGraph, Leiden
+  and Louvain must recover all eight (ARI >= 0.95, each population F1 >=
+  0.9), FlowSOM the seven common ones, and the rare one when asked for more
+  metaclusters than populations. The same checks must fail on data with no
+  structure and on random labels.
+- **macOS is tested on the public mirror.** A macOS job in ci.yml runs only
+  on openflo-public, where GitHub's macOS runners are free (they bill at ten
+  times Linux on the private repository), so each mirror refresh runs the
+  suite with the GUI on macOS.
+- **The PyPI release waits for the tests.** A version tag ran only the type
+  check before building and publishing; it now runs the whole CI workflow
+  (tests with the GUI, type check, privacy gate, install checks) first.
+- **A status message stays put after the gate tree rebuilds.** On Windows,
+  adding FMO gates (or an FMO refusal) could have its message replaced by the
+  selected gate's count, because the tree re-selecting its own row was taken
+  for a click.
+- **The editor opens on Tk 9.** The drag-and-drop package pinned at 0.4.3
+  stopped OpenFlo from starting on any Python that ships Tk 9 (uv's 3.14
+  and 3.15 builds, python.org's macOS installer from 3.14.5). It is now
+  0.6.4, and if drag-and-drop still can't load, the editor opens without it
+  instead of not at all.
+- **Identical groups keep p = 1 with SciPy 1.18.** SciPy 1.18 returns NaN
+  for a Mann-Whitney test where every value is the same, so a cluster or
+  population with the same share in both groups read NA in the comparison
+  CSV, differential abundance and group statistics. It is 1 again on every
+  SciPy.
+- **Cluster populations of a sample reopened without its clustering read
+  NA, not 0.** A session whose processed data was not found reloads that
+  sample from its FCS file, without the cluster column: every cluster read
+  0 events under a group holding all of them. They now read NA, and the
+  load says which results were not found.
+- **A session keeps its previous save.** Saving over a session keeps the
+  earlier version as `name.backup.flowsession`. A session that can't be read
+  says so in plain words and offers that backup, and resuming after a crash
+  skips a damaged autosave for the newest one that reads.
+- **Save and close updates the session you opened.** It used to write only
+  the autosave, so the file you opened was left as it was at your last
+  save.
+- **Problems found while loading are shown.** A renamed channel, an event
+  count that disagrees with the data, a spillover matrix that could not be
+  used, or a workspace population that was not imported went only to the
+  log or to a status line the next sample overwrote. They now appear
+  together in one "Notes from loading" window and are kept in the history.
+- **Damaged or unusual FCS files.** A parameter without `$PnN` is named
+  `Ch<n>`; one without `$PnR` takes the range its `$PnB` allows (as FlowJo
+  does); two channels with one name load as `FL1-A` and `FL1-A (2)`; and a
+  cut-short or non-FCS file says so in plain words instead of showing a
+  parser error.
+- **Spillover keywords that name markers.** A `$SPILLOVER` listing marker
+  names (`$PnS`) instead of detectors is matched to the detectors when each
+  marker belongs to exactly one, instead of leaving the sample
+  uncompensated. Values past the end of the matrix are reported.
+- **CSV files in common layouts.** Load CSV detects semicolon, tab and bar
+  separators and decimal commas, skips `#` comment lines at the top, keeps
+  the first event of a file with no header row (columns `Ch1`, `Ch2`, ...),
+  and refuses a file with no numeric column in plain words.
+- **Undo after compensation or a transform change.** Undo can't reverse
+  those, and it used to undo the gate change made before them without a
+  word. It now says it can't, and the next Undo goes on to the gate change.
+  Undo no longer brings back the gates of a removed sample.
+- **Exported FCS files keep their keywords.** Every keyword from the source
+  that does not describe the file's layout is carried over (`$SRC`, `$OP`,
+  tube and well names and more; the source's `$FIL` as
+  `OPENFLO_SOURCE_FIL`), and the matrix the values were compensated with is
+  recorded as `OPENFLO_APPLIED_MATRIX`. Keywords about compensation are not
+  carried, so no reader compensates the values twice.
+- **A session sample whose FCS file is missing is no longer deleted by the
+  next save.** Opening a session skipped a sample whose file had moved, and
+  the save or autosave after it wrote the session without it and its gates.
+  The sample is now kept offline and saved back unchanged, and after opening
+  OpenFlo offers to look for the missing files in a folder (and its
+  subfolders) and relinks them by file name.
+- **A session's saved data cut short no longer loads part of a sample.** The
+  per-sample data beside a session is written whole or not at all, and one
+  that is damaged or holds fewer rows than recorded is rebuilt from the
+  sample's FCS file, with a warning.
+- **Log-amplified and gain-adjusted FCS files load as intensities**
+  (behaviour change). Integer data with `$PnE` log amplification was read as
+  raw channel numbers and `$PnG` was ignored; both are now applied on load
+  and recorded in the history. A session holding such a file now reads it
+  as intensities. Float data is left as stored.
+- **A gate on a channel a sample lacks reads NA with the reason**
+  (behaviour change). It used to admit every event, so the population read
+  100%. It now admits none and Statistics, the report and the tree show NA
+  ("channel X is not in this sample").
+- **FlowJo AND / OR / NOT gates are imported from and written to .wsp.**
+  They were dropped on import without a word, with everything under them.
+  An operand that can't be found reads NA with a warning.
+- **Pipeline (beta) says when an embedding can't run.** TriMap was ticked by
+  default, and without the `[embed]` extra it was skipped with nothing but a
+  log line: the run ended "Done: 4 ok" with no TriMap output. PHATE, TriMap
+  and PaCMAP are now unticked and disabled in the panel when not installed
+  (the tooltip says how to install them), and a run that asks for one anyway,
+  or whose embedding fails or makes no plot, lists it in the run's summary.
+- **The highlighted sample is always one on the plot.** After loading
+  several files the first was plotted but the last to arrive was
+  highlighted, so a gate drawn with "all shown" went to both and
+  compensation went to a sample not on screen. Now the plotted sample is the
+  highlighted one after a load, highlighting an unticked sample ticks it,
+  clicking a ☑/☐ box no longer moves the highlight, and unticking the
+  highlighted sample moves the highlight to one still plotted.
+- **The Copy gates dialog shows its Copy button.** Its four buttons sat in
+  one row about 560 px wide inside a 420 px dialog, so Copy, packed last,
+  got no room and only appeared if you widened the window. Select all /
+  Deselect all and Copy / Cancel now sit on two rows of a 460 px dialog.
+- **Copying gates no longer stacks duplicates.** Copying appended every
+  gate, so a sample that already held it (copied before, or drawn there with
+  "all shown") got a second, identical-looking row. A gate the target
+  already has is now reused; the same population drawn differently (same
+  kind, axes and name, under the same parent, and only one such) takes the
+  copied geometry in place, keeping its id, on/off state and children; the
+  status line counts both. Anything less certain is appended as before.
+- **A FlowJo export names each unnamed gate the same in every sample.**
+  An unnamed gate was written under the editor's per-sample gate id, so
+  the same gate could be `g2` in one sample and `g3` in the next (any
+  sample that had once had a gate deleted numbered on from there). FlowJo
+  matches populations across samples by name, so its tables and group gates
+  didn't line up. An unnamed gate is now named for its axes (`FSC-A x
+  SSC-A`, or the channel of a 1-D gate), and siblings with the same name get
+  `(2)`, `(3)`. Before: 6-sample export, populations `g1`/`g2`/`g3` with
+  ctrl_d1 off by one. After: `FSC-A x SSC-A` / `FL1-A x FL2-A` in all six.
+- **Closing a large session no longer freezes OpenFlo.** "Save and close"
+  saved on the window thread, rewriting every sample's processed-data CSV:
+  about a minute for a large session, with every window marked "Not
+  Responding" and no message, and each extra click on ✕ queued another full
+  save. The save now runs in the background behind a "Saving session… n of
+  N samples" window, extra clicks only bring that window forward, and a
+  sample's CSV is rewritten only when its data changed since the last save,
+  which also makes the 5-minute autosave quick when nothing changed.
+- **Closing asks whether to save the session.** ✕ autosaved every time,
+  and a large session (every sample's data rewritten) froze every window
+  for up to a minute with no message, so the windows looked as if they
+  would not close; each further ✕ queued another save. It now asks: Save
+  and close, Close without saving, or Cancel. A second ✕ while closing is
+  ignored. (A stop-gap: making the save itself fast is its own change.)
+- **The update offer asks Update or Cancel.** It asked Yes / No / Cancel,
+  and No opened the release page: three ways to say yes. The release notes
+  link is now in the question's text.
+- **Export to FlowJo keeps gates on a channel of unknown scale.** A column
+  of an older OpenFlo CSV with no transform record has no transform to undo,
+  so its gates reached FlowJo in display units with only a log line (an
+  asinh cut at 4.0 became an intensity of 4.0). When the sample's raw FCS
+  can be read, the channel is now matched to it rank for rank, as FlowJo
+  will gate it, and the gate keeps its events (a cut exactly, a polygon to
+  within its re-traced edges). Without the FCS the gate is still written,
+  as stored. Either way a warning after the export names each such gate.
+- **A sample on the same transform with other settings is put on the
+  editor's.** A library export on asinh cofactor 5 loaded beside samples on
+  the editor's asinh (cofactor 150) kept cofactor 5: the same events plotted
+  at median 4.45 beside 1.16, its axis labels were wrong, and the Expression
+  window pooled both scales into one violin. The parameters are now compared,
+  not just the method, for the first sample too. A restored session's gates
+  move with the data: a cut drawn on cofactor 5 held 380 of 800 events, and
+  without converting it held none after reopening.
+- **A population that cannot be computed reads NA, not 0.** A gate whose
+  region is unknown (a FlowJo ellipse not converted, a boolean that lost an
+  operand, an ellipse whose covariance describes no ellipse) selects
+  nothing, and so does every gate under it or built on it. Statistics,
+  Frequencies and its comparisons, the report and the gate's status line
+  showed that as a measured 0 events. They now show NA with the reason,
+  Frequencies leaves it out as missing, and the tidy CSV writes NA.
+  - NOT or OR over such an ellipse now admits no events too; NOT over it
+    admitted every event.
+- **A FlowJo gate imported onto a channel of unknown scale reads NA.** Its
+  coordinates are in FlowJo's linear units and cannot be converted to a
+  channel whose display scale is unknown (an older OpenFlo CSV with no
+  transform record). It was kept as is, comparing a linear cut against
+  values on another scale, with only a log warning. It is now kept but
+  selects nothing, and its population, everything under it and every
+  boolean built on it read NA with the reason.
+- **`openflo-run`'s shared clustering and comparison files.**
+  - A run with one group removes an earlier two-group run's comparison
+    files (`compare_<A>_vs_<B>.csv`/`.png`, `compare_shared_clusters.csv`)
+    and says so, instead of leaving them as if it had made them. A file
+    that cannot be removed is named and the run exits 1.
+  - Channel labels match across samples despite common real spellings: a
+    fluor name at either end ('CD3 FL5' on FL3-A), Greek letters
+    ('CD8α' = 'CD8a') and detector names that differ only in case
+    ('FL7-A' / 'fl7-a'). Such channels were dropped before.
+  - Pooled and unpooled events follow one noise rule: each finite event
+    takes the label of its nearest labelled pooled event, so a sample's
+    "Unclustered (noise)" share no longer scales with how much of it was
+    pooled. Only rows with a non-finite marker value stay noise.
+  - The shared PhenoGraph's RAM check applies with the default
+    `--workers 1` too; with too little free RAM it runs on one core and
+    says why.
+- **The compensation editor's Apply keeps what was computed on the
+  sample.** Apply replaced the sample's data with a fresh load of its FCS.
+  Clusters, embeddings, U: abundances and _pos flags were lost, so their
+  populations selected no events and gates on them selected every event. A
+  downsampled sample came back with all its events.
+  - Apply now re-compensates the sample's own events in place: every event
+    and column stays, and each channel keeps its transform.
+  - The status line names what was computed before the change and asks
+    you to re-run it.
+  - A sample opened from a CSV with no record of its compensation and no
+    FCS behind it is refused, with the reason. It used to fail with "could
+    not parse FCS".
+- **A session with Leiden or FlowSOM populations can be saved.** Saving,
+  and the autosaves, failed with "Object of type int64 is not JSON
+  serializable".
+- **Gates that cannot be evaluated select nothing, and say so.**
+  - A boolean gate over a boolean that lost an operand (or over a gate
+    under one) now admits no events. With the inner boolean empty, NOT over
+    it admitted every event, and OR with it was the other operand alone.
+    Clear, Cut, session restore, .wsp ingest and template apply now also
+    name the booleans built on a broken one.
+  - An ellipse whose covariance is singular or not positive definite
+    admits no events instead of every event. Exported to FlowJo it is still
+    the all-pass rectangle, and the warning says FlowJo will count every
+    event.
+- **Open-sided gates are drawn in full on every view.** A quadrant, an
+  imported max-only range, or a polygon opened at an axis end drew only part
+  of its outline on a fluorescence channel's log and symlog views (no
+  outline at all on symlog). An open interval bound no longer draws a line
+  or stretches a linear view to ±1e12.
+- **Editing overlapping polygons grabs the nearest vertex.** A click on one
+  polygon's corner took the first-drawn polygon's corner ~9 px away; the
+  nearest vertex (then the nearest edge) of all polygons now wins.
+- **FMO gates are placed at the FMO's linear cutoff.** The FMO dialog took
+  the percentile in the FMO tube's own display coordinates and used it as a
+  gate value on the stained sample. An FMO on asinh cofactor 5 beside a
+  stained sample on cofactor 150 called 5.36% positive where 50.52% is
+  right; an FMO of unknown scale gave 0.0%.
+  - The cutoff is the FMO's percentile on its linear values, mapped into the
+    stained sample's display coordinates. With a parent gate selected, the
+    FMO is read through it in the stained sample's coordinates.
+  - An FMO or stained channel of unknown scale is refused, and the status
+    names the channel, the FMO and why.
+- **A logicle scale inferred from the value range is marked as a guess.**
+  A column of an older OpenFlo CSV with no transform record, whose values
+  all lie in the logicle range, was taken as logicle. A dim channel on asinh
+  falls in that range too: Statistics reported a linear median of 1793.94
+  where 80.56 is right.
+  - The entry is kept as logicle but marked guessed, through re-saves,
+    session restore, re-transforms and pooled Workspace runs.
+  - Its statistics are titled "(linear, guessed logicle)" in the Statistics
+    window and its CSV. Expression and Group comparison name the samples
+    whose medians rest on a guess.
+- **`marker_expression` names the samples it could not use.** A sample of
+  unknown scale went into `differential_test` as NaN and was dropped
+  without a word. The groups now list omitted and guessed samples, and each
+  result row carries `omitted_a/b` and `guessed_a/b`.
+- **Compensation-control names.** The rule took any word ending in "comp"
+  ("recomp", "precomp", "Decomp_1", "Autocomp", "TComp_1") as a compensation
+  control. It now names the bead brands written as one word (UltraComp,
+  OneComp, VersaComp, AbC, ArC). "Non-comp" and "No-comp" are samples.
+
 ## [2.6.2] - 2026-10-09
 
 ### Changed