chromosomix 0.5.0

What changed in chromosomix 0.5.0, released on 4 August 2026.

Added

• Right-click the map to mark the cell under the pointer. Vertical line, horizontal line, crosshair, loop — each row showing the coordinate it would use, so what is about to be marked is visible before it is — and copy coordinates. The two ways in that existed were typing a coordinate you already know and arming a mode and clicking; neither answers “this one, the one I am pointing at”.

• Lanes beside the map are draggable, and carry their own menu. The border resizes a track, the body reorders the stack, and dragging along the axis still pans the map — which it has to, since that is what dragging a track has always done. The direction of the first movement tells the three apart, so nothing needs a modifier key and no gesture was given up for another. A right click on a lane offers its colour, its range, hide and remove: the two controls a reader beside the map actually reaches for, where the track is rather than across the window in a card.

• Hand-drawn lines are continued into the lanes. A vertical rule is a column and is drawn under the map; a horizontal one is a row and is drawn beside it; a crosshair is both. Loop anchors are small triangles on each stack’s outer edge, pointing back at the map. A mark called out on the matrix is a coordinate, and the tracks beside it are what a coordinate is usually being compared against — the marks were on the map alone, which is where the comparison could not be made.

• Tick any number of genomes or sequences and see them as one map. Each on the diagonal and every block between them off it — the picture the pair view showed for exactly two, without the two diagonals that make an off-diagonal block readable. Both tables carry checkboxes now, with “select all” over whatever the filters leave visible (filter to completeness above 90, then take all of them) and “clear”; the button names two and counts more, because MAG_7 × MAG_12 → is worth more than a number and four names are longer than a button.

  It is a view state — view=selection&genomes=a,b,c, or regions= for sequences — so a selection is a link, a history entry and a back button, and genomes=a,b is precisely what view=pair used to be. The pair view stays where it is: the overview’s cross-genome clicks lead there and so do links already sent.

  The read is the whole-file scan with most of the file left out. A selection is scattered through the pixels, so there is no shorter span to walk — the gate says so, and the progress bar and cancel are the ones the whole-genome read just gained. The axis is built from the selection in file order rather than the order it was ticked in, because a matrix whose rows jump about is a picture of the clicking; it is pooled against its own size, so four MAGs are drawn at four MAGs’ resolution however large the file they came from is; and a genome resolves to its contigs, so ticking four MAGs and ticking their contigs are the same map. Checked on every fixture against the same oracle the pooled read uses: a scattered selection equals the dense block rolled up over the same grouping, cell for cell.

• The app says what it is on the way in. The mark — the same drawing as the favicon, the install prompt and the dock icon — is now a masthead on the landing page: mark and name on one line, centred, spanning 80% of the drop zone below them, with the tagline under and 20 px of air above. Both are scaled by one clamped font size, so the pair stays one thing from a phone to a wide window. In the header it sits at 26 px before the name, sharing its centre — the brand row was baseline-aligned, and an image’s baseline is its bottom edge, so the mark hung visibly high against the words. The first screen used to open on an instruction, which made it a dialog rather than a page. It is read from the 192 px file npm run icons:app already writes, not a fifth copy of the shapes.

• “whole genome →” in the genomes and chromosomes tables. The whole-file view was reachable from the chevron beside the root crumb and nowhere else, which is a corner of the window to find something the table in front of you is the question about. The metacool gets its scan, the cooler its map; the button is absent on a metacool’s per-genome contigs table, where “the whole file” is not what is being asked.

• The cooler’s whole-genome read has a progress bar and a cancel, the same two the whole-metagenome scan has always had — it is the same pass over every pixel of the same file, and the difference was in the plumbing rather than in the wait. What made it possible is the reader walking a span asynchronously: a cancel arrives as a message, and a synchronous loop never lets the message queue run however many microtasks it awaits, so a read that could take twenty seconds could not be called off and could not say how far along it was. The contig roll-up is watched the same way, since it walks the same pixels.

• Hand-drawn marks can be saved. The Manual annotations card grows the export row the Aesthetics panel has — a format, and a Save — writing lines as BED and loops as BEDPE, one base wide and half-open. Both come back through this app’s own sniffer and parsers as an overlay on the map, which is the property manual-export.test.ts checks rather than the text; and both are what bedtools reads, because a coordinate someone chose outlives the session they chose it in. Nothing to write is a disabled button rather than an empty file.

• The whole-genome view of a cooler is read at the resolution it is drawn at, and there is no longer a file size that defeats it. It used to ask for one cell per file bin, which at 24,228 bins is 2.3 GB of values in one array — and a browser serves at most 2 GiB in one, whatever the machine has, so a whole yeast genome at 500 bp was offered, computed for several seconds, and then refused. The picture it was refused for was one nothing could have shown: the renderer pools any raster to 4,096 px per axis, so those 24,228 rows were always going to be drawn as about 4,000 pixels.

  So the pooling moved to the read. Every stride consecutive bins of a chromosome are summed into one cell, in the worker, straight from the stored pixels — the block that arrives is the one that was always going to be drawn. Yeast at 500 bp is 4,047 cells and 66 MB instead of 24,228 and 2.3 GB; at 200 bp, where the dense block would have been 14.7 GB, it is the same 66 MB. A human cooler at 10 kb, 381.9 GB dense, is 67 MB and 760 kb cells. The read is bounded by construction at every bin size a file can have.

  What arrives is an ordinary block one bin size up, not a new kind of object: the stride restarts at every chromosome, so a cell is stride bins of one sequence and a run of cells is that sequence. The panel is told the coarser bin size and every coordinate, tick, tooltip, region box and figure caption is then right without knowing pooling happened — a box drawn on the 3 kb overview still opens that region at the file’s own 500 bp. It is the relationship an .mcool’s 5 kb level has to its 1 kb level, computed on the way through rather than stored, and the sums are identical to metacooler aggregate over the same squares (pooling.test.ts checks every fixture, cell for cell, against the dense read rolled up by the existing folded aggregation).

  What the view still costs is the pass over the stored pixels, which is what it always cost, and that is now what it asks about: above ten million pixels the map waits behind a button that states how many there are and what resolution they will be summed into. Below it nothing is asked. The whole-metagenome scan is gated for that same reason, and the two views now behave alike down to the wording.

• Both whole-file views are kept across navigation. They are the expensive reads, so the shell holds whichever one it built, keyed on what was read (uri, level, balance): leaving one for a chromosome and coming back returns to the picture rather than to the gate. Changing the level or the balance is a different read, so that one is rebuilt and the old block released rather than left in memory unread.

Removed

• Captions on hand-drawn marks. A text box under every mark, for a sentence the figure already makes. Gone from the card, from the model and from the drawing — and with it the label helper, the halo it needed and the type size it was written at.

Fixed

• A right-click on the map no longer navigates. mousedown fires for every button, and the map’s press/release pair had no guard, so a right-press was answered as a click on release: right-clicking a whole-genome view opened whatever cell was under the pointer, and the menu the press was for arrived over the view it had just left. The primary button only, now — and the same guard on the side panels’ drag handle, where a right-press armed a resize the browser’s own menu then sat on top of.

• Both ends of a loop are marked on both 1D axes. A loop is a statement about two loci, and a reader looking at the tracks under one end of it is asking what is at the other. Drawing the row anchor only beside the map and the column anchor only under it answered half of that on each edge, and neither half where the comparison is made.

• The masthead’s mark sits on the word’s centre line. Centring the two boxes left it visibly high: a line box carries the font’s ascent and descent, so the ink of a word sits below the middle of its box while a square’s ink is its box. Measured on the rendered pixels — 10.5 px of offset at a 97.6 px font — and corrected in em, so it holds at every size the clamp produces.

• A block too large for one typed array says which limit it hit. A browser serves at most 2 GiB in a single Float32Array whatever the machine has: measured in Chrome on a 16 GB Mac, 2,140,000,000 bytes is served, 2**31 - 4 is not, and four separate 1 GB arrays live side by side. It is a limit on one array, not on the tab, and the refusal used to blame “this machine”.

  It is now checked ahead of the memory budget and cannot be overridden by a confirmation, because there is nothing to override; the message names the largest square block that can be made (about 23,170 × 23,170 bins) and the ways through. The whole-genome view no longer reaches it at all — it reads pooled — so what it governs now is a region, a genome or a pair.

  A test asserted that yeast at 500 bp loads densely, and was passing, because it exercised the estimate and the estimate never tried to allocate anything.

• The Annotations panel scrolls as a whole, not card by card. Twelve little scrollbars where the reader wanted one. overflow-x: auto was set on the cards so a wide one could scroll sideways inside a narrow column — but a box with overflow-x: auto and overflow-y: visible is not a thing CSS has: the visible half computes to auto, so every card was a scroll container in both directions, and a flex item shrinks by default, so each one was squeezed and scrolled inside itself the moment the panel was full. flex: 0 0 auto on the cards is the whole fix. .side keeps the opposite behaviour deliberately: a view’s sidebar fills the column so its contig list can scroll where it is.

• Less white between a key and its value in the File card. The key column was a fixed 9rem, set when values were elided and had to start at a predictable place. They run to their full width now, so there is nothing to line up with, and the column takes what its longest key needs.

• More room before any of that. The memory budget is now half of what the device claims rather than a third, and the assumed memory where the browser says nothing — Firefox and Safari, on every machine — is 8 GiB rather than 4. A 20,000-bin block, 1.9 GB, was refused outright on both of those engines and now loads. Guessing high costs nothing since the allocation itself is caught and translated; guessing low was costing pictures the machine could draw.

• The whole-metagenome view is a view state, not a mode the shell holds. It was the only landing view of a file with no link, no history entry and no back button — pick it and the URL did not move, so it could not be shared, and reloading landed on the genomes table. It is view=overview now, a peer of the two tables it shares the root menu with, and everything that follows from that follows: the level and the colour scale are in the hash like every other view’s, the back button leaves it, and the plot re-measures when a side panel is dragged — which, being mounted outside the shell’s own view slot, it did not.

• The three panels resize — which, despite 0.4.0, is the first time. The drag handle, the remembered width and the double-click reset all shipped, and none of them could be reached: the handle was a sticky flex item at the end of a column that was already full, so it was laid out below the fold, and sticky only holds an element that scrolling would carry away — it never pulls one up into view. The negative margin meant to cancel its height could not, because a percentage margin resolves against the containing block’s width whatever axis it is on. So the strip was real, painted, and permanently out of sight.

  It is now positioned against the column, and the column no longer scrolls: its body does. The heading staying put while the cards move under it is the same change seen from the front.

Changed

• Saving a figure is one row: resolution, format, Save. It was a DPI dropdown, then a line of two buttons labelled PNG and PDF, two rows further down — which made the choice of file into the action itself and left the resolution stranded above the thing it applied to. Now the row reads the way the sentence does: 300 dpi, PDF, Save. Both defaults are unchanged; PDF simply stops being the second button and becomes the one that is already chosen.

• Cards in General scroll sideways rather than eliding or wrapping. A URI cut off at Hum66.met… and a genome name broken across three lines are the same failure: this panel is where you go to read the identity of what is on screen, and an identity you cannot read whole is not one. Keys and values now keep to a single line and the card scrolls under them. Hovering still gives the full text, for the reader who would rather not.