random-sketches

Logs

5 Nov 2026

Koi: music (楽, key n: 静 Stillness, sparse koto phrases over a drone; 流 Flow, a slow samurai-ambience piece with shakuhachi, koto figures, low strings and far-off taiko), all synthesised in the in scale on D. Rain now sounds like rain: a soft dark hiss and dull pats in place of the tinny sine drops.
Koi: rain (雨: Dry / Drizzle / Rain, key r, ?rain=): drops computed in the shader on a hashed grid, each a small ring, with faint high plips when sound is on. In ink mode, ripples and raindrops bend the ink and paper a little and are shaded like a pale wash. Technical write-up in BLOG.md.

3 Nov 2026

Koi: Ink mode after CCTV’s “Ink”: fish painted as keyframes (8 fps by default) on textured paper with feathered brush edges, and an ink-in-water simulation between them, so each fish sheds a light smoke trail shaped by its wake (tail jet flicked with each beat, shed vortices) that fades faster where the water moves
Koi: Ink is now the default style (Ink, Paint, Real), painted styles go down to 2, 1, ½ and ¼ fps, and the realistic pond has a choice of bottom (底: 石 Pebble in clear water, 空 a sky-blue pool, 黒 a dark liner) in place of the green algae.
Koi: the watercolour is pigment on textured cold-pressed paper (relief, fibres, pigment settling in the hollows), with wet-in-wet bleed and ragged blooms, subtler ripples, and a black-and-white ink style (写 / 絵 / 墨). Painted styles get 止 Still to hold a frame and 齣 to drop to 24 / 16 / 12 / 8 / 4 fps. The realistic view now simulates caustics every frame on the GPU from the wind waves and touch ripples (each bottom texel traced back to the surface, brightness from the Jacobian), and has 視 Top / 斜 2.5D / 巡 Orbit cameras, turned by a two-finger, right-mouse or shift drag. ONLY=shots node smoke.js passes.
Add Koi: a koi pond seen from above, for phones. One composite pass traces each pixel through the rippled surface: refraction onto a generated pebble bottom, photon-traced tileable caustics, a canopy-dappled sun, a blurred shadow map of fish, pads and leaves, and Beer-Lambert water colour along the bent path, with a Fresnel canopy reflection and sun glints. Koi are lofted bodies and translucent fins bent in the vertex shader along a 16-point path-following spine with a travelling wave. Each has its own wandering, depth and speed noise, swims by beat and glide, avoids the edges, keeps apart from neighbours and lines up loosely with them. There are 12 varieties with noise patterns in rest space, scale rows, eyes, metallic sheen and butterfly fins. Lily pads with water lilies, petals or leaves by season. Touch modes Ripple (rings, wakes; the fish shy away), Feed (pellets the fish rise to and gulp) and Call (the fish gather under a held finger). Tilt gives parallax, moves the glint and makes floaters drift downhill; shaking sloshes the water and scatters the fish. Animated watercolour finish (washes, edge pooling, granulation, boiling outlines at 8 fps). Synthesised sound with a shishi-odoshi, Wake Lock, adaptive render scale, offline PWA. node smoke.js 11 checks.

1 Nov 2026

Sail Trim: light-air pointing fixed. The genoa’s sheeting angle with the sheet in hard is now fitted (16 degrees, a toe-rail track; it was 8) and the calibration adds 36 pointing checks (sailing 7 degrees off ORC’s beat angle mustn’t beat ORC’s VMG); the steady-state solver now runs until speed settles (light air settled too slowly and read fast). VMG beat angle in 4 kn 48 degrees (ORC 46.6, was 36), 6 kn 46 (43.7). 2.6% RMS against the polar.
Sail Trim: Polar tab, the model’s auto-trimmed polar against the ORC certificate for the wind speeds you pick, with RMS error, VMG optimums and your boat; computed in Web Workers built from the model’s source (works from file://).
Sail Trim: timed windward-leeward races (race.js): 0.25 / 0.5 / 1 nm legs, 1 or 2 laps with a leeward gate, 1 / 3 / 5 min start with signals, OCS, a 360 for touching a mark, finish through the start line; race clock and next-mark HUD, Course tab with laylines and track, minimap, Time x1 to x8, ORC VMG target time and bests in localStorage. node --test 24 (with an autopilot that sails two laps), node smoke.js 16 checks.

1 Nov 2026

Add Committee Boat: race committee app. Fleet (import from a spreadsheet, sort A–Z or by sail number), starts per fleet (rolling, 5-4-1-0 or 3-2-1-0) with spoken signals, countdowns and the boats clear to cross at each gun; Crossed before the gun is OCS, after it a start, an OCS boat crossing again has returned; DNS / DNC, general recall. Finish: partial sail number or name puts a boat on deck, FINISH (on press) or a detected horn gives it the time, a time with nobody on deck waits for a name; edit h:m:s, DNF / RET, still racing and out lists. Results per fleet: PHRF time on distance or time on time, ties share points, codes score entries + 1, CSV / text. Clock from GPS fix timestamps or a manual offset (tap on the minute). Voice lookup (speech recognition), camera OCR (TextDetector or Tesseract.js) matched fuzzily to the fleet. node --test race.test.js (12), node smoke.js (32 checks).

31 Oct 2026

Sail Trim calibrated against a published ORC polar: the base boat is a Catalina 30 as its 2026 ORC Club certificate measures it (data.orc.org, 04560004IWP), and nine coefficients (sail lift and drag, spinnaker efficiency and clearance, hull friction, heel drag, wave-making resistance as a power of the Froude number, righting moment, windage) are fitted to its 117 speeds from 4 to 24 kn by calib/calibrate.js (auto-trim every point in parallel workers, Nelder-Mead on the coefficients, five rounds). 2.3% RMS error, 6% worst; VMG angles within 2 to 4 degrees from 6 kn up. New hull: Froude-number wave resistance, crew righting moment, windage. node --test model.test.js (16), node smoke.js (11 checks).
Add Sail Trim: sail trim simulator. Ten controls (mainsheet, traveler, vang, main halyard, cunningham, outhaul, backstay, jib sheet, jib car, jib halyard) shape three sections of each sail; each section’s angle of attack against a groove between luffing and stalling gives lift and drag; heel, hull drag with a wave wall, keel leeway and helm give boat speed. Plan view of the sections (twist) and an abeam view with draft stripes, jib luff and main leech telltales, luff wrinkles, mast bend and forestay sag; trim coach, auto-trim, gusts and shifts, tacks and gybes. node --test model.test.js (9), node smoke.js (6 checks).
Sail Trim: boat length slider 20 to 50 ft (areas with L², displacement with L³, righting moment with L⁴, hull speed with √L, so bigger boats are stiffer and faster); reefing (two slab reefs, lower head and centre of effort; auto-trim chooses the reef); spinnaker (sheet, guy / pole angle, pole height, luff curl, gets out from behind the main when squared back) and gennaker (sheet, tack line; blanketed by the main on a run), with broaches past 30 degrees of heel under a kite. New presets: gale beat with 2 reefs, gennaker broad reach, spinnaker run. node --test model.test.js (14), node smoke.js (10 checks).
Sail Trim: auto-trim on a jib run works. A stalled sail’s drag now falls off past 90 degrees to the wind (it was flat from 90 to 180, so a search started from sheeted-in sails found nothing to improve and stayed put), and off the wind the search also starts from eased sheets. Wing on wing: the jib poled out to windward, out of the main’s wind shadow, about 0.3 kn faster dead downwind; auto-trim chooses it, the coach suggests it and warns when it is too high for it. node --test model.test.js (15), node smoke.js (11 checks).

29 Oct 2026

Aquarelle: QuickShape is off by default (Preferences; existing prefs turned off once) and undo after it gives back the stroke as you drew it (one more undo removes it, redo goes forward). Palm rejection: touches are ignored while the pen is down or was used in the last 300 ms (hover counts), and touches already down when the pen lands are dropped, a stroke a palm began just before taken back (Preferences -> Ignore the palm).
Aquarelle fixes: undo of a stroke over most of the sheet failed and lost the step (it needed a second copy of the whole sheet in the undo atlas; then every later undo stopped working too) - tiles now swap in place through a one-tile scratch. QuickShape no longer turns shading back and forth into a wet ellipse ring (which a wet brush filled into a blot) when the pen rests: a loop must go once round its middle, a line must be straight, anything else stays as drawn; the hold is 650 ms.

1 Nov 2026

CarPlay Gravity: cracks no longer freeze the browser as they land. A new crack is drawn into the crack maps on its own, its glow blurred once as a layer, instead of redrawing every crack with a blur per line segment (in software rendering a crack went from about 4.3 s to 0.1 s to appear, a spider web from 8.7 s to 0.13 s). The glass shader runs at 1.5x at most on high-DPR screens and skips the ink blot and LCD streaks where they can’t show, about half the GPU time a frame with one crack. node smoke.js 44 checks.

28 Oct 2026

CarPlay Gravity: Dashboard slabs split along the crack lines: a crack cuts the slabs still in place where the glass cracked, breaking across a bend too, and the break lines between slabs are uneven and slanted instead of a grid. The physics engine takes convex polygons. A live panel is drawn once a frame and shared by its slabs. node smoke.js 44 checks.
CarPlay Gravity: Dashboard panels break apart into slabs (map 3 x 3, cards 2 or 3) that look whole at home, burst apart when knocked loose and fall back to 74% size so the pile has room to tumble, each still drawing its live part of the panel; Restore reassembles them. Moving glass takes turns at the live blur fairly. Triple-tap cracks vary more in angle (up to about 40 degrees off level). node smoke.js 42 checks.
CarPlay Gravity: a slow triple tap leaves a single hairline (no blot, no streaks, a light shock); a quadruple tap turns the crack into a spider web of radial cracks and rings. Tap timing uses the event time, so a slow frame no longer drops a tap. node smoke.js 41 checks.
CarPlay Gravity: cracks need a triple tap, and how quick the three taps are sets how hard the knock is (more and longer cracks, more LCD streaks, a bigger shock); thicker crack lines; a smaller ink blot at the impact. node smoke.js 39 checks.
ScoreShift: a C section stays in C after a page in D. A sharp or flat hugging a staff’s first note is that note’s, not key ink, and a system of staves none of which prints a key is a keyless section like two single staves in a row (Sibelius 2 viola pp. 2-4: staves opening on F♯ or C♯ and the divisi systems read D major). Fixture scan_sibelius2_va_keyless.png; bun test 55, eval unchanged.
CarPlay Gravity: cracks run across the screen instead of out like a star (one to four roughly sideways lines through the impact, fewer forks and rings), and a broken LCD shows thin blue and red light streaks, mostly columns, some rows, appearing as the crack spreads; they replace the stuck pixel column. node smoke.js 38 checks.
CarPlay Gravity: Cracks strength slider in the sheet (0.25× to 3×, default 1.8×, ?cracks=) for crack count, reach, shard shift and shock; cracks run out from the impact over half a second with a glint on the tips; each wedge between cracks shifts and turns as one piece and shades as it tilts (after ksenia-k’s broken glass pen); smaller colour split on text; Shake and Drop moved clear of the Direction label. node smoke.js 38 checks.
Aquarelle: fitting the page (100%, the % button, opening a page) uses the whole screen below the top bars; the tool strip, sliders, palette and zoom float over the page instead of shrinking it.
CarPlay Gravity: ripples off by default (a crack no longer splashes when they are off), bolder crack lines with a dark shadow edge, and the WebGL pass holds GPU memory only while it runs: the GL canvas and textures shrink to 1 × 1 when idle, crack maps at up to 1.5 px per logical px with their 2D canvases freed after upload, the frame texture updated in place. node smoke.js 36 checks.
CarPlay Gravity: cracks and liquid ripples as a WebGL2 after-effect (fx.js) over the canvas frame. Shards shift the picture under them, crack edges split colour and catch the light, the panel bleeds ink from the impact with an oily rim; touches send ripple rings, drags leave a wake, Shake and cracks splash. Ripples On / Off in the sheet (?ripples=off); 2D fallback without WebGL2. Smoke Chromium runs with SwiftShader WebGL. node smoke.js 35 checks.
Aquarelle: one-tap Hide all panels (eye at the top right; a faint eye stays in the corner to bring them back; Tab and the four-finger tap do the same).
CarPlay Gravity cracks look like a cracked screen: a few long, nearly straight cracks in uneven directions with kinks and forks instead of a spider web, plus panel damage (an ink blot at the impact with a coloured fringe, colour bleeding along the cracks, a chromatic split at their edges, a stuck pixel column).
CarPlay Gravity: the sidebar’s glass pill is a falling piece too (it was baked into the backdrop). Its clock, recent apps and buttons pass through it until they have come apart once (World.ignorePair), then collide like any other pair; it tips over its foot when it drops. node smoke.js 32 checks.
CarPlay Gravity: Liquid Glass blur on the canvas (each glass piece blurs and saturates what is under it with ctx.filter; pieces at rest blur the backdrop once, cut to their outline, pieces in flight the live frame, four a frame), Light appearance with light wallpapers, Clear icons, iOS 26’s widgets page (calendar, Home accessory tiles that toggle), an Apple Maps-style 3D driving view (extruded buildings, route ribbon, labels on the ground, fog), double-tap cracks the glass and knocks pieces loose, a new Phone glyph, Helvetica after SF Pro in the font stack. Gravity starts after 5 s or at the first touch, 1.5 g by default, Shake twice as hard with a screen quake. 1.6 ms CPU a frame headless. node smoke.js 32 checks.
CarPlay Gravity looks like CarPlay in iOS 26 / 27 (from Apple’s WWDC25 and iOS 27 screenshots): floating glass sidebar with the Dashboard / app-grid button, squircle icons painted in canvas with a lit rim in Default and Dark appearances, labels on glass chips, CarPlay’s default app order (4 × 2 on 800 × 480), page dots, layered Celosia-style wallpapers (Indigo, Rose, Teal) and Ocean; the Dashboard is a map with an arrival pill and glass cards for the next turn, Now Playing and Up Next. The HTML controls are gone: an options sheet drawn on the canvas (sidebar button, Settings, or M) with segmented controls, a strength slider, a direction dial, Shake and Drop / Restore, over the frosted running scene. node smoke.js 24 checks.
CarPlay Gravity on one canvas, laid out for 800 × 480. Pieces are specs that draw themselves (cached bitmaps keyed on what they show; maps and gauges live, the gauge face cached) instead of DOM elements, taps hit-tested through each body’s rotation; layouts in logical pixels with the height fixed at 480. Small screens drop the bezel and fold the controls into a ☰ drawer. 0.7 ms CPU per frame headless. node smoke.js 19 checks (both sizes).
Add CarPlay Gravity: a CarPlay-style screen (Home grid, Dashboard, instrument Cluster) whose pieces fall, after mrdoob’s Google Gravity. Each .phys element is measured in the live layout and then moved by a CSS transform from its rigid body, so it keeps working while it tumbles (map driving, music, clocks, gauges, taps). Gravity: Down with a dial / arrow keys, device Tilt, Drive (lateral g from a simulated route’s turns, rumble, potholes), Zero-G; 0–3 g. Pinned pieces are knocked loose by a hit over 260 px/s; Restore eases them home. physics.js: boxes (Box2D Lite clipping) and circles, warm-started sequential impulses, soft mouse joint. node --test (5), node smoke.js (13 checks).
Aquarelle panels can be moved and folded: a grip on the palette, pen bar, tool strip and size / water / paint column; drag to place (kept relative to the nearest screen edges, so it survives rotation and resizing; a panel away from the edges floats and gives the page its room), tap to fold to the current colour / nib / tool / value, tap the folded panel to unfold, double-tap or Preferences → Panels to put back. node smoke.js 163 checks.

26 Oct 2026

Squash Court: storage meter and clean-up for recordings. The meter shows the recordings’ size (analysis and video) against the browser’s quota, with Keep (persistent storage). Each recording shows its size and can drop its video only. Clean up removes video or whole recordings by age (a day, a week, a month, all but the newest 3). When the quota runs low, the video of the oldest recordings is dropped automatically (an option, on by default), never the analysis or the live session; if that is not enough, video recording stops. pickCleanup in session.js, tested.
Squash Court: record video with camera sessions (option) and a rally list. The camera’s picture is recorded with MediaRecorder in 10 s segments (each a whole, seekable file, saved to IndexedDB as it ends with the camera fitted on it; a lens switch starts the next one); a recording plays its video on the stage with the players, ball and court drawn over it, in step with the 3D replay. Rallies: a recording’s rallies with time, length, shots, who hit how many and the last shot; a click plays that rally and stops at its end. On a 25 s camera session from the Crouin v Coll file: 3 segments, 4.5 MB, 5 rallies. bun test (67).
ScoreShift exports with pen annotations. When a page has marks, ⬇ PNG and ⬇ Transposed PDF ask first: draw them or leave them out, and which colours (say a teacher’s red marks but not your own blue ones). The choice is remembered. PDF pages are drawn in the export workers like before; each worker rebuilds the page’s music and draws its marks at the page’s own scale, on the stretched staves of the transposed page. Annotations now follow those stretched staves in the Transposed view on screen too (before, a mark could sit a little off its note where a key signature got wider). inkexport.js; 4 unit tests and a smoke check that renders the exported PDF and finds the red mark and not the blue one (49 checks).
ScoreShift ✎ Annotate: write on the photographed music with a pen (or a mouse; a finger writes until a pen has been used, then fingers scroll, and two fingers always scroll). The ink sticks to the music under it: a fingering to its note, a circle to the passage it surrounds, a margin note to its moment on the staff. It stays put through corrections, the Original / Transposed views and reopening the project. Strokes written close together in quick succession become one mark (a “3” in two strokes). Four colours, three widths, an eraser (and the pen’s eraser end), undo / redo, clear page; tap a mark to recolour or delete it, or to re-attach it when a correction deleted its note (drawn dashed). Saved per page in the autosave and in the .scoreshift project, and drawn into the PNG export. 2 new smoke checks (48 in all).
notation-ir pen annotations (annotations.js): ink attached to the music, not the page. A stroke is anchored to the note under it (fingering, breath mark, word), to a span of notes on one staff (circle, bracket: it stretches with the notes, its margins stay a fixed size), or to a moment on the staff (margin notes). It is stored in staff spaces relative to that anchor and redrawn on any surface: it keeps its note through corrections, on the transposed page, at another scale, and re-engraved onto other systems (a span splits at the line break). A mark whose note is deleted is drawn where the note was (or at its moment), flagged lost, and can be re-attached. Surfaces now report each event’s own staff top, so a mark on the bass staff of a grand staff is measured from that staff. 18 model tests on a toy surface, plus the minuet photo (correction, transposition, deletion) and StaffInk’s real layout (re-engraved narrower); 122 notation-ir tests, smoke unchanged (33 + 46).
notation-ir surfaces (surface.js): one interface for anywhere music is visible - lines(), place(id, edge) (a list: a note can show in several places, or none), eventAt(point), view(), optional barOf / barAt / placeAnchor / lostAt. Marker layout and hit testing, selection bands, playhead maps and time-from-a-point are written once against it. StaffInk’s engraved page (staffink/surface.js) and ScoreShift’s photo (scoreshift/surface.js) now implement it, and both apps’ hand-written geometry is gone (app code −97 lines net); no change in behaviour, all 33 + 46 smoke checks pass as before. 100 notation-ir tests, including the generic features against a toy surface and both real surfaces.
Squash Court Live / Analyse and recordings. Live: the picture alone with a HUD (time, shots, bounces, distance per player; tap for Analyse), the panels not drawn; Analyse: the panels, below the picture on a phone (beside it on a phone on its side), the less used controls under Options; a phone opens the camera in Live. Recordings (session.js): a camera session saves itself to IndexedDB (every 20 s, on hide, on stop), Save this session for any other; Open shows it as a match with a 3D replay (slider, play), Download / Open… as .json. The live clock now carries on across a lens switch (a new stream restarts at 0, which reset the speed chart). Fixed the whole page scrolling on a phone (the screen-reader zone list escaped the side panel). Video / still-frame recording is in TODO.md. bun test (63), node smoke.js (all ok).
Squash Court camera zoom: a 0.5× / 1× / 2× / 3× / 5× bar in camera mode (camzoom.js), from the track’s zoom range where the browser exposes one (Chrome on Android, Safari 17+), else the phone’s other back lenses (Safari’s Ultra Wide and Telephoto by name; unnamed Android lenses as Lens 2, 3 to try). Switching lens keeps the video element and the heatmaps; the court is found again. Remembered per browser. bun test (61), node smoke.js (fake phone camera: steps, zoom applied, lens switch).
Squash Court ball: heard impacts help pick the ball among rival flights (a flight scores for a heard hit just before its first frame, a heard wall or floor just after its last, a heard fitted impact). The page sends its heard impacts to the ball worker, tools/match.mjs passes each chunk’s own. Measured on two files: with good sound kept flights start on a heard hit 2.3× as often as chance, changing ~4% of kept flights; on a highlights soundtrack without the hits, nothing. Earlier today: long ball tracks fitted piece by piece instead of rejected whole at their first bounce, flights entirely next to a player dropped, one ball at a time (green-box-on-flight 1–16% → 5–67% on cc.mp4); older flights fade on the video and the 3D view. bun test (55), node smoke.js (all ok).
ScoreShift: a part that is only on some of the page’s systems (an orchestral part going divisi on two braced staves for a few lines) starts its bars with the bars they stand under (score.js alignParts, measure.at, used by timeline() and the notation-ir adapter), instead of from the top of the page, where it played over the opening bars (Sibelius 2 viola part, 8 of 17 pages; found through MidiViz #046). Fixture scan_sibelius2_va_divisi.png; 54 tests, fixtures unchanged.

25 Oct 2026

Squash Court response patterns: for one player’s shots (all of them or one type), where the opponent took the ball in reply, over 12 zones. Each zone shows its share, with the busiest shaded; hover, tap or the arrow keys list a zone’s reply shots (“Volley Cross Drive”). Replies are graded High / Medium / Low from the shot and where it was taken. On a 41-minute game: 30 and 19 replies located. bun test (54), node smoke.js (22).
Squash Court coaching panel (coaching.js): for each player, where they stood when the opponent struck the ball, and how often that was on the T (within 1 m); volleys, where they were taken and where the opponent was then; and every shot’s first floor bounce on its own map by shot type. Racket hits are found by carrying the ball back to a player’s reach, and hitters are labelled through each rally by alternation. Three maps (both players or one) plus a summary; also in match.json, the export and the match.mjs summary. On a 41-minute game: hitter known for 115 of 415 shots, so the numbers are thin until the ball is fitted on more shots. bun test (51), node smoke.js (21).
Squash Court: Predicted path toggle (video and 3D view), and the prediction fades as the ball passes it: the part already passed by the video’s time is drawn faint, all of it fades out over 0.6 s after its end. node smoke.js (all ok).
Squash Court: the predicted ball path is drawn over the video too (dashed orange, rings at its impacts, red where it goes out or into the tin); a colour legend under the 3D view; a camera view box to show or hide the camera’s field of view there. node smoke.js (all ok).
Squash Court: a player benchmark on real frames, and T domination. tools/squashai_fetch.mjs and tools/pbox_eval.mjs score player detection against the SquashAI datasets’ 2,392 hand-drawn player boxes on 1,253 PSA broadcast stills. Baseline: court fitted on 83% of frames; on those, 33% of players found with feet in place, 47% overlapping one of ours, feet 12% of box height off, 33 cm on the floor. It is broken down by video, by distance and by overlapping pairs. T domination per player (tcontrol.js): of the time both are seen, the share each is nearer the T, plus time within 1 m of it and the average distance. Shown in the heatmap titles and the readout, and written to the export, match.json and the match.mjs summary. bun test (45), node smoke.js (20).
Squash Court ball path in the 3D view: fitted flights with their floor shadow, the ball now with a drop line, and the latest flight predicted onward, dashed (predictPath: drag, gravity and the simulator’s bounce model through walls and floor, stopping at the second bounce, an out line or the tin; impacts as dots, out in red). A drive predicted from its first 0.18 s stays within 3 cm of the truth through the front wall and two bounces. bun test (38), node smoke.js (19).
Squash Court whole-match identities across camera cuts (tools/match.mjs). Each track’s shirt colour is the median of its clean reads (not while the players are within 0.9 m or occluded); tracks seen together in one camera stretch are linked into chains labelled as a whole, so a restarted track keeps its partner’s opposite; k-means centres refined under those constraints. On a 7½-min PSA highlights file: 63 of 63 tracks with their own shirt (was 40 of 65, 1125 detections under the wrong player), detections on the right shirt 0.61 → 0.99. --merge <out> re-merges saved chunks. bun test (37).
Squash Court player identities. Players that overlap in the image are no longer swapped: the second person is searched for on visible pixels only (the first one’s box counts as hidden), each detection carries a shirt colour read from the chest band of its measured height (it was read from a 1.75 m box and caught heads), and PlayerTracker scores both pairings jointly with position and shirt, refuses a box on a clearly different shirt, and coasts a hidden player for up to 1.5 s instead of starting a new track. On the main camera of a 7½-min PSA highlights file: identity switches 136 → 5, steps on the right shirt 0.74 → 0.99, steps with only one player found 451 → 356 (shirt-labelled, so the count favours the colour term; six close moments checked by eye). Synthetic players unchanged (median 1.7 cm). bun test (37), node smoke.js (19).
Squash Court: player names, captions, time slider, plan the right way round. Players panel: type the names (saved per video file, ⇄ to swap), used in the stage labels, readout, heatmap titles and the commentary (lines rebuilt from their shots, so renaming rewrites the log). Captions: the newest commentary line over the video for a few seconds, the feed keeps the log. Time slider under the video; a seek restarts the ball and player tracking. Fixed the floor plan, front wall, heatmaps and match.png being mirrored (x = 0 is the right wall facing the front, and was drawn on the left). node smoke.js (19).
Squash Court commentary and a wider panel. Live shot commentary (commentary.js): each shot built around its front-wall contact (measured, or a flight carried on to / back from the wall) and named lob / drop / kill / boast / drive, straight or crosscourt, with length, tightness to the side wall and pace; corner sprints and a player pinned in a corner; lines rewritten in place when a flight is refitted, Speak to read them aloud, shots in match.json from tools/match.mjs (main camera only: flights from court-level and side cameras are left out). Side panel resizable (drag its edge), two columns: commentary feed, floor plan, front wall with every ball contact and the last one in green (last floor bounce too), each player’s own heatmap with distance and top speed. Clear now also resets heatmaps, bounces, speeds and commentary. bun test (36), node smoke.js (17).
Squash Court: who hit it, sound, speed charts and whole matches. Bounce dots in the hitter’s colour (the last front-going flight carried back to its racket sound, nearer player there). Impact sounds (audio.js: sharp onsets above 2 kHz) decoded from the opened video; a heard impact where a lost or not-yet-started flight would meet a surface becomes a bounce, and the audio delay is estimated from seen impacts only when it beats chance. Player speed chart (last 30 s). tools/match.mjs: a whole match file in parallel chunks (court tracked through cuts, players, ball, sound), players linked across cuts by shirt colour, per-player heatmap / distance / speed over time and every bounce with its hitter, match.json (opens in the page) and match.png; a 15-min PSA file in ~3 min on 30 processes. Player fixes found on whole files: the walk past a score bug must give a person’s height (the bug’s own bars were taken for feet, 20 % of detections at the back wall), caption-tab edges may have small gaps, the bars under a bug are masked, foreground that never moves within a camera stretch is masked, and the tin band seen level is no longer taken for a graphic. bun test (30), node smoke.js (15).
Squash Court ball bounces: floor bounces on the floor plan and front-wall hits on the rectified front wall, a dot per bounce and a heat glow (25 cm cells, blurred) where they cluster, Bounces toggle, counts in the readout (and how many landed behind the short line), in the JSON export. Bounces come from fitted impacts and from flights that end on the floor, kept for the whole clip (bounces.js). The ball worker now gets every new video frame through requestVideoFrameCallback instead of only the frames the court worker analysed, which had left gaps of several frames. bun test (26), node smoke.js (13).
Squash Court speeds and heatmaps. Player speed from a 0.6 s least-squares slope of the tracked position, top speed and distance covered; the ball’s speed from the fitted flight’s velocity at the current frame, in m/s and km/h on the frame and in the readout. Heatmap toggle: time per 25 cm floor cell per player, blurred, in the player’s colour on the floor plan; both in the JSON export. Feet below a score bug are now found through it (a lunging player in front of the bug was placed 0.9 m off from the head), caption tabs on a score bug are found and the graphic margin is smaller, and coasting tracks stay in the court. Ball fitting 4× cheaper (RK2 at 1/100 s, a quadratic smoothness test before any 3D fit), so the ball worker keeps up in the page. node smoke.js (12).
Squash Court videos play with their sound: Sound on / off button next to Pause (falls back to muted where the browser refuses unmuted playback). node smoke.js (11).
Squash Court players and ball in metres. Players: the court’s surfaces as background (robust colour model per region, lines and seams masked), occupancy search for 0.6 × 1.75 m people over the floor, feet from the bottom outline, score bugs masked and feet hidden behind one placed from the head, two identities in video; synthetic median 1.7 cm with the true camera. Ball: three-frame difference over distinct frames (25 fps broadcast in 50 fps files), ball-sized isolated blobs, 2D tracks cut where one drag ballistic flight stops fitting, 3D flights fitted through the camera with impacts labelled front / floor / side / back / racket; synthetic median 2.4 cm, 90th percentile 9 cm (14 cm cut automatically; a frame-timing bug in the rally simulator had inflated it to 21 / 44 cm). On PSA footage flights are found on wide views but loosely. Page: player rings with x, y, flights and impacts on the frame, floor plan, front wall and 3D view, readout lines, in the JSON export; players and ball in two extra workers so camera tracking keeps its rate. rally.js, rally-worker.js, tools/rally_run.mjs; bun test (23), node smoke.js (11).
Aquarelle: pastel, infinite pages, compact controls. Soft pastel and Pastel pencil (new ink-pass mode 7): dry, opaque, chalky, deposited only where the paper’s height beats a pressure threshold, so light strokes catch the grain and hard ones fill it; covers dark paint, tilt broadens, Smudge blends. Infinite page (Gallery → + → Infinite page / canvas): a sheet-sized window over unlimited paper in 128 px tiles; Sim.shift() moves every texture, the wet state and the undo history, changed tiles leaving are read back, gzipped into a new IndexedDB tiles store (v3), and tiles near the window are decompressed ahead so panning never waits; ruling follows world coordinates. Size / water / paint as three value chips over one slider (tap to choose, drag a chip to change it), zoom by dragging the % between + and − (the zoom bar is half as tall), a layout pass that keeps every panel on any screen (⋯ overflow menu for the top-right bar, scrolling tool strip, shrinking slider, popovers clamped and scrolling; checked at 320 × 568, 560 × 420, 844 × 390, 1280 × 600), tool names on tap or hover, + in the page bar for a new page (N). node smoke.js 118 checks.
Squash Court video tracking without the line map (track.js). Model line samples are projected with the last camera and a colour profile is read across each at the frame’s own resolution (peak of the both-sides-alike response nearest the prediction, in the learnt colour of that kind of line, continuing along it), so a frame costs 1–7 ms instead of ≈ 50 ms. A camera that has not moved (judged by the joint Gauss-Newton step, < 0.2 px) is held exactly, and offsets averaged over held frames refine it every 12 frames; pan / tilt / zoom are fitted on a fixed mount (position prior), which removes the zoom / distance drift, and the mount is released on cumulative evidence that it moves (then a constant-velocity prior). The full tracker covers lost or doubtful frames, the first frame after a slow detection and long gaps; cameras held for a while are remembered and picked up again after a cut without a detection. Fixed-mount synthetic pan / zoom clip: position within 1.6 cm on every frame, floor 0.4 cm at the 95th percentile (full tracker 3.9 m); moving-camera clip 10 cm (2.4 m); PSA footage 90–99 % of frames held, zero jitter. tools/track_eval.mjs, tools/make_ptz_video.mjs, ?track=full, ?mount=moving; bun test (18), node smoke.js (9).
Squash Court on real PSA footage. Seeds from the service boxes (box sides, backs and short line as fitted lines, corners by intersection, floor homography decomposed into a camera over a ladder of focal lengths); per-line scoring (each line by the fraction matched, floor lines most, one model line per image pixel, direction must agree), the focal / distance valley probed, the tin fitted at 0.48 or 0.43 m (PSA men), reverse-angle cameras above the front wall allowed, distortion no longer folds far-off points into the frame. Fixes the user’s Silicon Valley Open frame, which fitted the front wall but slid the floor sideways onto glass-panel seams. Confidence needs three matched lines including a floor line, so end cards and logos are rejected. Tracking refines tight first and probes the zoom valley; the page re-detects on a cut. 138 of 169 frames sampled from two PSA rally compilations detected; tools/batch.mjs; bun test (14), node smoke.js (9).
Add Squash Court - the camera from a squash court’s lines. A pixel is on a painted line when the pixels either side are alike and both differ from it the same way (three scales, four directions, normal from the doubled-angle sum), so lines of any colour on wood, dark floors, white walls and glass fire but edges don’t; sub-pixel centre lines linked into straight segments, short ones (text, grain, crowd) dropped, then a nearest-line distance map. ~6000 seed cameras from where squash cameras stand, scored by model-on-image coverage, local stochastic search, then Gauss-Newton ICP over rotation, position, focal and k1 (Tukey, shrinking scale, plausibility bounds; the back-wall line ignored from behind it), judged both ways at 640 and 480 px. Video tracks from the previous frame and re-detects after three bad frames. UI: overlay, click to metres on floor or walls, rectified floor and front wall, 3D view with the frustum, camera JSON. Synthetic sweep 17/20 (median 1.6 cm, 0.03°, 0.08 % focal); 11 of 18 PSA broadcast stills fitted by eye (misses are close-ups); detection ≈ 450 ms, tracking ≈ 60 ms per frame. bun test (11), node smoke.js (9: demo pose, click, export, tracked clip at ~19 fps headless).
Aquarelle notebooks, zoom and a two-finger lock. Any painting can have pages (Gallery → + → Notebook lined / grid / dotted, or Canvas → Add a page): ‹ page › at the top, › on the last page adds one, Page Up / Down, a page list with thumbnails to jump, add, move and delete. One page is in the simulation at a time; turning dries and saves it and keeps your zoom and column at the top of the next. Pages are stored per key with the notebook’s card listing them (older paintings are one-page notebooks), reopening at the last page; duplicate and delete cover every page. Zoom buttons (+ / % / −) at the bottom right; 100% is the whole page in the space the toolbars leave, pages open at 125% on 7” e-readers and phones (Prefs → Open pages at: Page, 125%, 150%, Width), starting at the top left; panning keeps part of the sheet on screen. Two-finger gestures are classified from their first ~10 px and locked: moving together only pans (no zoom, no rotation), spreading zooms, only a clear twist rotates. Fixed fit() reading the frame timestamp as a zoom when called from requestAnimationFrame. 90 smoke checks.
Sorairo puddle reflections and a darker trunk. Puddles now mirror the hill’s own things: the utility poles and their crossbars, the trunk and a lobed canopy in its season’s colour that thins as the tree goes bare, traced analytically in the terrain shader along the mirrored view ray (no second render) and wobbled by the drops. The trunk no longer glows pale from below: its branch tubes were wound inside-out, so back-face culling showed the sunlit inside of the far wall, and the bloom threshold is raised (0.92 to 1.35, radius 0.4) so sky and bright cloud no longer bloom over the tree. Smoke: 27 shots.
Sorairo rainbow after the rain: as rain clears with the sun out and below about 40°, a soft painted bow round the antisolar point (primary 40.7–42.3° violet to red, fainter reversed secondary at 50–53°, brighter sky inside, darker Alexander’s band), fading over about 8 s or a day and a half of a time-lapse; rainbow slider / ?rainbow=. Puddles reflect a varied overcast sky with drifting lighter and darker masses, stronger ripples, and a low puddle camera shot to see them up close. Smoke: 27 shots.
StaffInk ink A/B toggle: Settings → Ink (Auto / Low latency / Standard) swaps the ink canvas between a desynchronized and a standard 2D context (a context’s options are fixed at creation, so a fresh canvas element replaces the old one) and shows which one the browser actually granted. The mode chip names the active ink while comparing; ?lowlatency=1/0 still overrides. Auto stays standard on Android. Smoke: 32 (toggle, writing after a swap, Android UA picks standard); the offline check now finds the current cache instead of a hard-coded version.

26 Oct 2026

Clips can be reordered in StaffInk and ScoreShift. The clips appear in playing order as coloured chips (StaffInk: top right; ScoreShift: next to the play-mode switch). Tap a chip to edit its clip, drag it sideways to play it earlier or later, or use ⇠ earlier / later ⇢ in the clip’s toolbar (Alt+←/→ in StaffInk). A clip keeps its colour when it moves (the colour follows its letter), and the first clip always starts on its own. Shared moveClip (arrangement.js) and cliporder.js.
ScoreShift finger selection in ✂ Markers: hold still for a moment, then drag across notes; the page stays put while you select and scrolls on near the edges. A quick swipe still scrolls the page. Long-press menus are suppressed on the page while Markers is on. 90 notation-ir tests; smoke StaffInk 33, ScoreShift 46.

24 Oct 2026

Sorairo rainy days and puddles. Four seasons and the time zoom bring their own rainy spells (Animate → rainy days, ?weather=clear for none): a day or three at a time, deterministic in the day number so scrubbing replays the weather, about one day in five most of the year, more in April, about half the days in the tsuyu rainy season, a third in the September autumn rains, none while it snows; the time zoom keeps the real sky for its real-time start. Puddles collect in the hollows of the path as it rains, ringed with darker wet mud, reflecting the grey sky (more at grazing angles), with spreading raindrop ripples. Smoke: 25 shots.
Playhead lines in StaffInk and ScoreShift: a line across the staff glides through the music while it plays, so the position shows through rests and long notes and across jumps. It follows line breaks and the tilt of a photographed staff, and the view scrolls to keep it in sight. Programs get one line per cursor, coloured like the clip it is playing (a canon or layered clips show each voice); normal Play gets one per part. Shared notation-ir playheads.js (timeMap, cursorsAt, headsAt, drawPlayheads), and perform() trace segments now carry their clip marker. ScoreShift draws the line on its own canvas, so a frame redraws only the line. 88 notation-ir tests; smoke: StaffInk 30, ScoreShift 44.
Sorairo day cycle runs round the clock (noon, afternoon, golden hour, magic hour, night, dawn, morning, noon) over 60 s and loops without a cut. Rain (Scene → rain, ?rain=0…1, ?weather=rain, rain.js) at any time of day and in every animation: a low stratus deck with darker drifting rain bands, colours drained toward a cool grey at the same brightness, no direct sun, shafts, flare or cloud shadows, raised ambient and thicker haze, wind-slanted stateless rain streaks that give way to snow, darkened wet grass and path, no birds, airliner or fireflies, and rain hiss, patter and drips in the sound. Smoke: 23 shots (day-cycle dawn, two rain frames).
Sorairo dawn and morning keyframes, so mornings no longer replay the evening backwards. Dawn: a cool lilac sky over a pastel peach horizon, rose cloud edges, the last stars and low valley mist (keyframes now set fog density too); morning: crisp, cooler light as the mist burns off. The time-of-day slider runs once round the clock (0 noon … 4 night, 5 dawn, 6 morning, 7 noon) with the moon setting and the sun rising; the time zoom uses it (night 20:00 to 04:00, dawn about 05:25, morning about 06:50), and the higurashi call again at dawn. Smoke: dawn and morning shots (20 in all).
Sorairo time zoom (?anim=zoom, journey.js): starts at the real clock and date in real time, then over 44 s speeds up to one day every 8 s (the sun crosses the sky east to west), one day every 3 s (about four days and nights), and one month every 4 s, where it holds. The rate is eased in log steps and the simulated time is its integral, so the progress slider and ?p= jump to any point. The hour runs the full 24 hours through the time-of-day keyframes; past about a day per second the days lock to one hour (15:00 default, lock), each night shrinking to a blink, and from then on every day gets its own clouds and light. The date sets the season (blossom about 1 April, first snow mid December). Clock overlay with date, time and speed; Start time zoom from now. Also a slow camera drift during every time-lapse (motion-control-style swing around the tree with a slight push and rise, drift, pauses while dragging). Smoke: 18 shots, all three animations checked live, and Chromium is found in any Playwright browser folder.
ScoreShift marker authoring on the photo (✂ Markers). Tap a bar to select it, Shift+tap or ⇢ Extend to grow the selection, or drag across notes with a mouse. Play or ⟳ Loop plays just that passage, with speed and metronome, so you can practise bars 12–16 at ×0.75 with clicks. ＋ Clip keeps it as a bracket on the page (repeat, octave, delay, ⧉ with previous); ⤳ Jump continues at a note you tap. A switch in the transport sets what Play plays: the page, the page taking its jumps, or the clips. A marker that lost its note to a correction can be re-attached to the nearest note or to one you tap. Programs autosave per page and go into the project. The arrangement logic moved to notation-ir arrangement.js, shared with StaffInk; clicksFor gives metronome clicks for any program (looped bars click every pass). 81 notation-ir tests, 43 ScoreShift smoke checks.
ScoreShift key signatures from MidiViz (#046, its Sibelius viola part): opening signatures that start with naturals (a key change at the page turn) read as a cancellation (naturals alone, C) or as the new key (naturals then sharps or flats), each glyph keeping its type so the transposed page samples a natural as a natural; a natural needs 3 spaces of clearance to count as part of the signature. analyzeAfter(norm, fifths) reads a page in the key the previous page ended in: staves that print no key take it (and key changes inside a staff count from it), while a printed signature, a cancellation, a keyless section (now marked key.keylessSection) or a printed key that disagrees keep the page’s reading; analyze(norm, { fifths }) is now that default rather than an override, and the page’s own inference still runs. PDF renders cancel with an AbortSignal and clean up their page; a reader pool that fails to open closes the readers it had opened. Fixtures spliced from the Telemann crops (tools/make_key_fixtures.mjs); fixtures and degraded conditions unchanged (pitch 98.3 / 95.3 / 96.5 / 90.4%). 51 tests.
ScoreShift: a flat whose bowl a ledger line cuts from its stem is re-paired (the stem keeps the stub of the bowl’s bottom, up to 0.6 spaces wide); the bowl alone read as a double sharp, so B♭ became B𝄪 (Sibelius 2 viola part, p. 4, found through MidiViz #046). New fixture scan_sibelius2_va_p4.png; fixtures and degraded conditions unchanged. 52 tests.
ScoreShift: an accidental read half a step off its note (a flat touching a hollow head loses the bottom of its bowl to that note’s stem) is attached to the head a step away when it hugs it and its position is within 0.7 of a step: bar 198’s E♭5 half note in the Sibelius 2 viola part no longer reads E5 (and the D♭ of the chord in bar 196). Pitch on scan / photo / phone fixtures 95.3 → 95.4, 96.5 → 96.8, 90.4 → 90.5%; clean unchanged. New fixture scan_sibelius2_va_p4_top.png. 53 tests.

23 Oct 2026

Sorairo time-lapse animations (Animate folder, ?anim=day|year). Day cycle: noon, afternoon, golden hour, magic hour and night over a 50 s loop, holding the night and fading through dark back to noon. Four seasons, a year over 80 s from seasons.js (sampleSeason(s), s = 0 deep winter): patchy snow melt (one snowAt field shared by ground, grass and flowers, mountain snow line climbing), the bare tree blossoming pink and dropping petals, spring green deepening to summer, wildflowers, the autumn turn through all six colours to red, shedding in turn order (per-card drop rank, cards shrink away as the canopy goes bare while the leaf fall thickens), then a bare crown of forking twigs with snow on the limbs, falling snow (snow.js, stateless points in a camera-centred box), buried grass with straw tips, greyer winter skies and quieter birds and insects. Clouds run on their own clock (uCloudT) 30× faster while playing, and each day of the year gets its own cloud cover and light, for a time-lapse look. Controls: play, scrubber, jump to season, seconds per loop, cloud speed-up; dur, p, play=0, season, lapse. 16 smoke shots plus a live check that both animations play.
StaffInk selection and playback markers. In Select, dragging across notes with the pen or mouse selects them like text (with a finger: hold, then drag; Shift+click and Shift+←/→ extend). The selection covers every staff in that stretch of time, and Play plays just it (⟳ Loop repeats it). ＋ Clip keeps it as a marker: repeat ×n, octave up or down, delay in beats, or play it together with the previous clip (⧉, for canons and layers). ⤳ Jump adds a jump from the end of the selection to a note you tap. With markers, Play plays the arrangement: the score taking its jumps, or the clips in order (switch at the top right). Markers live in the score (autosave, undo) and follow edits; erasing a marker’s note removes the marker. arrange.js, 7 tests.
Shared marker drawing in notation-ir (markers.js): clips as coloured brackets split at line breaks and stacked when they overlap, loop counts, jumps as arrows, label chips, hit testing, and lost markers red and dashed where their note was. ScoreShift draws the page’s saved program on the photo; after a correction deletes one of its notes, that marker turns red and Play names it. 6 tests; smoke checks for both apps (StaffInk 29, ScoreShift 42).

25 Oct 2026

Aquarelle: back to the first interface’s floating toolbars, keeping everything since. One button per brush on the left (tap the selected one again for its settings; smudge / eraser for their shape), size / water / paint sliders beside them with values on top, the pigment palette with the big mix well and + Mix along the bottom (rainbow well or mix well for the colour wheel), menu, gallery, wet dot and Pen

Paint top left, undo / redo / hair dryer (hold, double-tap) / tilt / layers / share / full screen / help top right. The sheet is fitted into the space the panels leave, so nothing sits on it (on narrow screens the sliders move to the right). Pen mode swaps the palette for the pen bar. E-paper theme redone for the new panels; phones drop tilt, full screen and help into the menu. 65 smoke checks.

22 Oct 2026

notation-ir: ScoreShift ids keyed to the page (ss:L<staff>:q<x in quarter spaces>:<voice>, barlines ssb:L<staff>:q<x>) instead of the order events are met, so a deleted or added note no longer shifts every later id. Anchors match a head re-read a few pixels off (within half a staff space, same voice first) and report a deleted one (AnchorError) instead of choosing its neighbour; checkProgram lists every marker that lost its note. ScoreShift keeps the page’s program in the .scoreshift project and playProgram() replays it after reopening, with the corrections applied. 69 tests, 41 ScoreShift smoke checks.
notation-ir writes values by beat. A meter tree (bar, beat groups, beats, subdivisions; simple, compound and additive meters such as 7/8 as 2+2+3) decides each value: a span that is a whole node, or starts and ends on its child boundaries, is one value; a dotted value whose dot reaches the middle of the next unit is kept (h. from beat 1 of 4/4, q. on beat 2 of 3/4); the standard syncopations (q h q, e q e, s e s) are kept, others tie at the beat, so the middle of a 4/4 bar and the two beats of 6/8 stay visible; rests take exact beats only; a gap joins the rest next to it (a bar of silence is one bar rest). The input’s own values still win where they fit (written), and beats / syncopation switch the rules off. Ignoring the printed values, it reproduces the engraved rhythm of 14 of 15 fixtures. 65 tests.
Sorairo sound, all synthesised in Web Audio (audio.js, no samples), off until a gesture: ♪ button, M key, Sound folder with music / ambience volumes, ?sound=1. Ambience follows the scene: two-band wind noise scaled by the wind setting with gusts and leaf rustle, songbirds by day (more when the flock crosses), a far jet rumble under the contrail, higurashi cicadas at dusk, crickets at night. Music: slow generative felt piano over a detuned-saw pad, 64 bpm in D major on the royal-road progression (IVmaj7 V iii7 vi, ii7 V7 I), pentatonic melody landing on chord tones, lower and sparser at night; generated stereo hall and a limiter. Also since the first version: bird flocks and a contrail that cross wherever the camera looks, an airier canopy with a canopy gaps slider (airy), and a mixed autumn palette: each leaf pad one of six colours (green to orange-red and brown) with warm shadows, moved through early to late autumn by a turned slider (red-dominant at 1); the old all-orange autumn kept as maple. node smoke.js: 12 shots plus an offline render of the sound for day, dusk and night (fails on silence, NaN or clipping) and a live ♪ button check.
Aquarelle (both interfaces): paint landed below the pen on Android. The canvas was styled 100vw × 100vh but its buffer and the pen-to-paper mapping used innerWidth × innerHeight; on Chrome for Android 100vh is the height with the address bar hidden, so with the bar showing the picture was stretched (offset 26 px at the top of a 7” screen, 84 px at the bottom, reproduced headless by making the box taller than the window). The canvas is now sized and mapped from its measured box (ResizeObserver, visualViewport), and a resize mid-stroke waits for the pen to lift before refitting. Full screen button, and full screen on first touch (default on Android). v1/ gets both; it stays the classic toolbar. 65 smoke checks, run on both interfaces.
Aquarelle: the first interface (commit 08ecf21) kept at vibes/aquarelle/v1 for side-by-side testing, with its own settings key, IndexedDB database and offline cache so the two installs never touch each other’s paintings (the current app upgrades the shared database to version 2, which a version-1 open refuses, and each service worker deletes other caches under its own prefix).
Add notation-ir - a measure-free timeline IR shared by StaffInk and ScoreShift. Content is events on exact rational time (any tuplet ratio, no tick rounding) in voices, links (ties, slurs, hairpins) and directions (clef, key, meter, printed barline, dynamic); fragments carry their own time for ideas sketched off to the side; competing readings share an alt group. interpret() derives bars from the meter or printed barlines (printed ones win and re-phase the meter), splits notes across bars into ties, infers tuplets from the duration (1/12 = triplet eighth, 1/20 = quintuplet sixteenth), fills gaps with rests or leaves paper, and spells with a chosen accidental scope (bar / note / time window for free time); bar status (ok, pickup, end, under, over, free) is reported, never repaired. Layout as one least-squares fit with minimum gaps (pool-adjacent-violators): engraved, proportional or anchored to the ink / the photo. Anchors (note, printed barline, nth bar, point on a staff line) resolve when used, so they follow edits; positionAt maps time back to the page. perform() plays authored programs - clips, walks taking jump markers, seq / par / loop / seeded choose, gap compression, confidence floors, alternative readings per loop pass - on parallel cursors, with StaffInk’s articulation rules (slur legato, staccato, staccatissimo, tenuto, marcato, accents, fermata). Pitch, duration and the Web Audio player moved into it and now back both apps. Adapters: StaffInk <-> IR (plays exactly what StaffInk plays) and ScoreShift -> IR (same pitches at the same times on all 15 fixtures); ScoreShift’s ✎ Edit in StaffInk opens the page, as written or transposed, in StaffInk with a report of what it cannot hold (second voice, 5-tuplets, tenor clef). playProgram() in both apps. PLAN.md: model decisions, adapter contract, gaps, next steps. 57 tests; StaffInk 54 + 28 smoke checks, ScoreShift 45 + 40.

Aquarelle pen mode and e-paper. Pen

Paint switch: ink goes straight into the active layer through its own pass (no water, anti-aliased edge, even line within a stroke, undoable, composited only under the stroke): ballpoint, fineliner, brush pen, highlighter (transparent glaze), eraser nib, stylus eraser end; ten inks chosen to stay distinct on Kaleido colour e-paper plus six highlighter colours in a bottom pen bar, + / round button for the colour disc, touch-and-hold to pick from the page; notebook pages (1264 × 1680) with lined, grid or dotted ruling drawn in the screen pass. E-paper theme (white, black outlines, no shadows or animation, drying redrawn a few times a second), auto-detected for Bigme / Boox / Kobo / PocketBook; laid out and checked at the Bigme B7 Pro’s 632 × 840 CSS px. No desynchronized canvas on Android. Switching paintings now waits for a save in flight (a quick New could lose the previous canvas). 53 smoke checks.

Aquarelle, Procreate-style: gallery of paintings in IndexedDB (thumbnails, new canvas sizes, rename / duplicate / delete, import a photo as a glaze layer via per-pixel KM inversion); top bar with Actions (share PNG / JPEG / time-lapse, canvas paper / texture / colour / flip, prefs, help) and Water & drying, paint / smudge / erase with brush libraries and a brush studio, layers, colour disc; sidebar with size, eyedropper, opacity, water and undo / redo. Layers (add, hide, opacity, rename, clear, duplicate, merge, move, delete, all undoable in the stroke history) composited with Kubelka-Munk, one K + S texture each with shared scratch ping-pong for the active one; smudge (wet and dried paint); Flat brush; QuickShape (held stroke is taken back with its water and repainted as a line or ellipse); quick-pinch fit, four-finger tap hides the UI; pressure curve, right-hand interface; time-lapse frames read back asynchronously, replay and MP4/WebM export. Fixed blits after creating a framebuffer reading from the wrong one (silently copied the wrong texture), and an unused sampler bound to the render target dropping layer composites. 42 smoke checks.
Aquarelle faster and on plain white paper. Brush stamps go down in their own passes scissored to each batch’s footprint (all of a frame’s stamps at once, no extra substeps); the KM composite is written to an sRGB texture only over the wet rect, so the screen pass is one texture read plus a precomputed paper-relief tile (render 112 → 24 ms, substep 174 → 122 ms on SwiftShader at iPad size); fewer substeps per frame when frames run long (drying scales, flow stays at its stable per-substep rate); autosave reads back through pixel-pack buffers + fence and Float16Array (no main-thread stall); panels no longer use backdrop-filter. White paper and no visible texture by default (Menu → Show paper texture; the grain still drives the physics).

21 Oct 2026

Add Aquarelle - watercolour painting on a WebGL2 fluid simulation, scissored to the wet part of the sheet (async 48×48 wet-map readback grows/shrinks it, so dry paper costs nothing). Per 180 Hz substep: virtual-pipe shallow water (O’Brien & Hodgins / Mei et al.) with contact-line pinning that favours the paper’s valleys and capillary creep onto damp paper; pigment advected with the exact fraction of moving water; rim-weighted evaporation for edge darkening; Curtis et al. 1997 adsorption/desorption with granulation and staining. Pigment carried as Kubelka-Munk K/S amounts (specified as unit layer over white/black, inverted in linear light), so mixing on the brush, on the paper and glazing are subtractive. Tileable procedural papers drive lighting, flow, granulation and dry brush. Pointer Events with coalesced samples, pressure, altitude/azimuth tilt, palm rejection by pointer type, pinch/pan/rotate, two- and three-finger tap undo/redo, long-press colour pick; tile-atlas undo on the GPU that also rewinds what the water did; hair dryer, board tilt / device tilt; IndexedDB autosave of half-float pigment (gzip), PNG export, offline PWA. Surveyed Rebelle, Expresii/MoXi, Fresco Live Brushes, ArtRage, MignonSketch, Procreate and the literature (Curtis 1997, Van Laerhoven 2005, Chu & Tai 2005, Mixbox 2021, MNPR). bun test (10), node smoke.js (26 checks: real pen/touch/mouse strokes, flow, drying, mixing, glaze, undo/redo, lift, pinch, export, save/restore, iPad/phone layouts, offline) plus a painted landscape.
Add Line Rider 3D - Line Rider’s Verlet sledder (sled corners + rope, butt, shoulder, hands, feet; bindings break past 13% stretch in one 1/240 s substep, about a 4-unit drop) riding one-sided ribbons. Strokes are drawn on a sketch plane: a wall plane is 2D Line Rider seen from the side (drawing direction sets the riding side, Shift flips), turned 15° with A/D or continuously while drawing; a floor plane draws top-down at a grade. Upward-facing parts bank into turns for the speed expected from the drop since the flag. Line ends snap and move the plane there; the line tool bridges two ends anywhere. 3D needed two additions to the 2D model: runner grip (sideways slip turned into speed along the ribbon) and yaw alignment. Demo: drop, corkscrew loop, lip + ballistic landing, banked 180° turn, accelerate strip, swells. bun test (9) and node smoke.js (shots, demo ride, mouse-drawn strokes, snap, erase, steer).
Line Rider 3D: 2.5D side view is the default (G toggles 3D). A 22° camera looks at the sketch plane slightly from above and behind, pans instead of orbiting, turns with the plane, and while riding follows the rider and swings to keep the motion across the screen. Fixed menus closing as soon as they opened (selects blurred themselves on click). Smoke test finds whichever Playwright Chromium is installed and checks menu focus.
Add Sorairo - anime-style landscape in three.js 0.160 (no build step). Hand-wired post chain so passes can read several inputs: scene to an MSAA HDR target whose alpha carries sun occlusion (sky writes cloud density), radial light shafts from that mask, UnrealBloomPass, lens flare gated on sun visibility, luminance-only Reinhard so colours stay saturated, lift/gain grade, then an anisotropic Kuwahara filter (structure tensor, Gaussian-blurred, 8-sector polynomial weights after Kyprianidis et al.) with edge pigment pooling, canvas weave, vignette and grain. All materials are custom cel shaders sharing one set of lighting uniforms: two-tone terminator, blue-violet ambient shadows, a tree-only shadow map (billboards face the sun camera in the shadow pass), moving cloud shadows and sun-tinted aerial perspective. Five time-of-day keyframes interpolated by one slider. node smoke.js renders 7 still shots through headless SwiftShader and fails on shader errors or blank frames.

6 Oct 2026

ScoreShift saving and a CODA snapshot test. Corrections autosave in IndexedDB per file (content hash) and page and come back when the same file is opened again; 💾 Save project writes a .scoreshift file (source file + edits of every page + settings) that Open restores anywhere. Edits are anchored where symbols were recognised (edits.js): pitch steps, accidentals, values / dots / triplets / ties / graces, articulations, deleted notes and rests, clefs, keys, key changes, time signatures. snapshots/coda.json snapshots the whole CODA reading bar by bar; bun test lists the bars that change, UPDATE_SNAPSHOTS=1 accepts. 43 tests, 36 smoke checks.

3 Oct 2026

ScoreShift: appoggiaturas told from acciaccaturas by the slash (fixtures 6/6, CODA graces slashed), played on the beat for half the main note and exported as <grace/>, switchable in the popup. Transposed page: ties and slurs move with their notes when found whole (CODA 26 of 27), articulations and bowing marks move with their note, symbol ink lost under staff lines is put back when pasting (stems no longer break, hollow heads keep their edges more often), and no more white squares where a line runs through a moved head. 41 tests, 34 smoke checks.
ScoreShift bowing marks and grace notes on the transposed page. Up-bows and down-bows read by template + shape above the chord (read first so stacked staccatos/accents beyond them are kept, and taken back from text rows); shown, editable, exported as <technical>; bowing fixture 22/22, CODA bowings match the page. Grace notes move with the transposition (smoke check: all 7 grace heads ink at their new place). 41 tests, 34 smoke checks.
ScoreShift evaluated on a second IMSLP work: Telemann’s 12 Fantasias for solo violin (MusicXML #183793, 11,534 notes) against the typeset PDFs from it and two independent scans (Bärenreiter 1955, Kallmeyer c.1927). Pitch 98.2% typeset, 95.2 → 95.6% Bärenreiter, 86.0% Kallmeyer; keys 287/289, 253 → 256/259, 265 → 267/301. Fixes: ring closed by a sharp + staff line + next symbol is not a half note, flat bowls not hollow heads, digit loops after a key glyph ignored, no key inferred for a run of keyless staves with no key-like ink. Evaluator compares notes per column low to high (two voices) and keys before any key change.
ScoreShift grace notes: a second, finer head pass finds small heads; kept only with their own short stem and a flag, beam or slash, clear of other notes’ stems and chords. The CODA waltz acciaccaturas (bars 56, 58: G♯4 before F♯4) now read, take no time, play just before the beat, and export as <grace slash="yes"/>. New grace fixture (6/8 found); Telemann extras 22 → 26 (the new ones are real grace notes its answer key omits). 40 tests.
ScoreShift practice transport: playback position slider (start at any bar, Stop keeps the place), speed ×0.25–×2 applied live, and a metronome that follows each bar’s meter (compound meters by dotted quarter, pickups counted back from the bar line). The player now schedules notes and clicks a little ahead of the audio clock, so speed and metronome change while playing. Also: slurs followed through staff-line and barline cuts, hairpins from the widening gap between their strokes. 39 tests, 33 smoke checks.
ScoreShift CODA corrections + dynamics, hairpins, slurs, debug labels. Ties the user found missing on the CODA sheet: E5 bar 3 into bar 4 (a tie fused with the phrase slur, now found by following a thin curve between same-pitch heads) and G♯ bar 15 into 16 (removed last round by a wrong “portato” rule; Audiveris had it right, and the reference was corrected), so bar 16 now starts on G♯. Dynamics read by template against SMuFL and italic-serif renderings of every dynamic and common expression words (tools/make_dynamics.mjs), including ones touching stems; on CODA all dynamics are read and attached correctly. Hairpins and slurs recognised; dynamics drive velocity, hairpins ramp it, slurs play legato; MusicXML <dynamics>, <wedge>, <slur>. Interpreted view gets a Show menu toggling bar numbers, values, articulations, ties, slurs, dynamics/hairpins, voices and text boxes. New expr fixture. 38 tests.
ScoreShift: key changes are correctable in the staff popup (change, remove, or add one after any barline; pitches and the transposition follow). Bold scanned sharps whose crossbars fill the gap between the stems are split at the two most-inked columns: Bärenreiter scan keys 262 → 263/264, semitone errors 54 → 47, pitch 98.7 → 98.8%; fixtures 95.4 → 95.6%. Tried and dropped: reading accidentals that touch their notehead with the glyph classifier (false hits on beams). 37 tests, smoke +2.
ScoreShift reads key changes inside a staff (Telemann Fantasia 12’s maggiore / minore): accidentals after a barline at signature positions, cancelling naturals, plausibility rules from false positives on 48 IMSLP pages (lone glyph needs a double bar, naturals must cancel the whole key, text above the staff excluded). Pitches read per key segment, transposed per segment with the change redrawn for the target key, next staff of a part inherits it, MusicXML <key> per measure. Thick / repeat barlines on scans accepted. Telemann IMSLP: typeset 99.6 → 99.7%, Bärenreiter scan 98.6 → 98.7%, keys 261 → 262/264; two crop regression tests (33).
ScoreShift two voices, articulations, tie accidentals, text rejection. Bars read as two voices only when they show them (opposite stems or stemless against stemmed in one column, displaced rests) and only if that explains the bar better than one voice; each voice checked and repaired on its own; heads take their own stem instead of one running through another voice’s head; hollow heads touching a filled head still read as rings; MusicXML voices with <backup>. Staccato, tenuto and accents read (15/15, 7/7, 5/5 clean, no false marks), played (shorter / louder), exported and editable. Tied notes keep their accidental across barlines and system breaks. Rows of letters outside the staff are text, so “o”/”p” of tempo marks and dynamics are no longer notes. Double flats reassembled (5/5 clean). Two new fixtures (two voices; articulations + double accidentals); bar accuracy now compares onsets, so voice splits don’t matter. Clean: 137/140 bars; CODA still 67/67; Telemann unchanged. oemer’s notehead U-Net compared and written up (helps phone photos, too big and slow to adopt). 34 tests.

2 Oct 2026

ScoreShift evaluated on real IMSLP files with MusicXML ground truth: Telemann’s 12 Fantasias (MusicXML #236786 vs the typeset PDF #236783 and the scanned 1955 Bärenreiter #96616). tools/imslp_fetch.mjs (resolves IMSLP downloads in headless Chromium, renders pages), tools/groundtruth.js, tools/imslp_eval.js (edit-distance alignment on pitch, per-staff clef/key). Pitch 98.1 → 99.6% typeset, 96.0 → 98.6% scan; keys 227 → 263 and 243 → 261 of 264 (the recognition fixes landed in d384f84).
ScoreShift competing readings + a learned glyph classifier, evaluated against Audiveris. Audiveris 5.11 built from source (aarch64) and run on the same images; tools/compare.mjs scores MusicXML against truth or each other. Neural options surveyed (Legato 430 MB, SMT 86 MB: no symbol positions; homr AGPL; oemer segmentation for later): instead a 98k-parameter MLP (glyphnet.js, 130 KB int8, ~0.1 ms/glyph, plain JS) trained on 120 synthetic Verovio pages in 5 fonts (tools/make_glyphset.mjs, tools/glyphnet.mjs), reading note accidentals 29/30 on phone photos where the stroke rules read 20/30. Rules and classifier now compete per symbol (classifier may add or retype at ≥ 90%, never delete); split double sharps reassembled; rest values scored by template + classifier with runner-ups kept for the bar check; bar repair can drop a stranded unbeamed head; beam stacks stop at a slur; staccato dots no longer dot a rest. Pitch: scan 94.6 → 96.0%, photo 96.4 → 97.2, phone 83.0 → 91.6; bars on phone 89 → 97 of 124. CODA sheet now matches the adjudicated reference in all 67 bars (Audiveris 55). 31 tests.
ScoreShift reads rhythm: playback, MusicXML and MIDI. rhythm.js counts beams and flags per stem (beams stacked from the stem end, stopping at a slur), dots (rejecting tie pieces cut up by staff removal), grace notes, and matches rests, time-signature digits, C / cut C and tuplet numbers against SMuFL glyphs of 5 fonts rasterized by a new canvas-free path rasterizer; time signatures are read as a region split at the middle line (the digits fuse through it) with a prior on real meters; ties link same-position heads. score.js builds parts and measures and checks every bar against its meter: pickups and section ends may be short, otherwise the cheapest unambiguous edit (dot, rest value, unmarked beamed triplet, beam count, flag, half/quarter, false rest, ignored tuplet number) makes it full, else the bar is flagged. Interpreted view shows values and bar checks; tap to set value/dot/triplet/tie/grace, delete a rest, set a time signature or play from a bar. ▶ Play with a playhead at a set tempo; ⬇ MusicXML 4.0 (as written or transposed with <transpose>) and ⬇ MIDI. Fixtures gain every bar’s events and three rhythm tunes written as MEI (Verovio’s ABC importer drops quarter triplets and chord lengths): clean 99.6% note values, 123/124 bars exact; photo 96.0% / 111. Real page: the CODA viola audition sheet (Brahms 2 + Fledermaus, Finale-style engraving) reads all 67 bars adding up after fixing wide tilted half-note holes, a tempo letter read as a flat, a cut C read as a sharp, keys of different pieces on one page voted together, slur-fused barlines, a double sharp read as a whole note. Pitch on fixtures also up: clean 96.0 → 99.0%, scan 91.8 → 94.6%. 30 tests, 32 smoke checks.

1 Oct 2026

ScoreShift clefs on scans, real-scan regression test, faster PDF opening. Clefs touching the system bracket or opening barline were swallowed into a system-sized component: long vertical runs of oversized components are cut in the clef zone. A staff found twice (beam taken as its top line) keeps the version most slices support. Key votes break ties toward the longer signature. Haydn scan, 132 staves: clefs 127 → 131, keys 125 → 128. Three one-system crops of the scan (~45 KB each, 16 grey levels) with true clefs/keys as bun test cases (21). PDFs render on a pool of up to 4 pdf.js documents that keep their last page decoded; probe renders double as thumbnails; probes and export share a worker pool: 30-page scan opens in 6 s (was 12), exports in 22 s (was 27).
ScoreShift faster PDF export and better key signatures on scans. Export runs on a pool of up to 4 workers (recognise + rewrite + JPEG with OffscreenCanvas; main thread only paints PDF pages) and probes each page first so text pages skip analysis: 30-page Haydn scan 77 s → 27 s. Keys on 8 real score pages 98/132 → 125/132 staves: split sharp halves re-paired, sharp crossings rejected as noteheads, C clef curls joined to the clef, per-system then per-page plurality vote for unread or short-read keys (never across sharp/flat). Fixture pitch on simulated scans 93.1 → 93.9%. 24 smoke checks.
ScoreShift PDFs (IMSLP, Internet Archive): pdf.js 6.3 legacy build vendored with its JBIG2 / OpenJPEG wasm decoders (scanned PDFs are JBIG2 + JPEG 2000 layers), lazy-loaded and precached for offline. Opening probes pages at low resolution for staves and starts at the first page mostly covered by music (skips cover, title, preface music examples); thumbnail strip, per-page corrections, and “Transposed PDF” writes the whole file back as a new PDF of the same page size via a ~40-line writer. Paste masks are now cropped per symbol: a 350-note score page renders in 0.75 s instead of 8.8 s. Tested on a 30-page Haydn op. 17 scan: opens at page 11 in ~12 s, exports in ~80 s. 23 smoke checks.
Add ScoreShift - photograph sheet music, get the same page back transposed for clarinet, sax, horn or any of 18 instruments, edited in place: every pitched symbol is cut out of the photo and pasted back on its new line or space, key signatures and accidentals are rewritten (page’s own glyphs where it has them), ledger lines regenerated, paper and staff lines under the cuts refilled from the photo, and the line squeezed per system when the new key needs room. Layout, beams, slurs, dynamics and lyrics stay put, so rhythm and engraving never have to be solved. Surveyed SmartScore, PhotoScore/NotateMe, capella-scan, ScanScore, SharpEye, Audiveris, PlayScore 2, Soundslice’s ML scanner and Newzik Maestria (all re-engrave; none edits the photo in place) and the literature: Rebelo et al. 2012 and Calvo-Zaragoza/Hajič/Pacha 2020 surveys, Fujinaga run-lengths, Dalitz et al. staff removal, CRNN/CTC, TrOMR, SMT, Legato (2025) and Legato 2, oemer and homr for phone photos.
ScoreShift OMR: classical staff-first pipeline in plain JS in a worker (Sauvola + Wolf–Jolion binarisation, run-length staff metrics, sheared-histogram deskew, sliced five-line detection scored by completeness so ledger stacks lose, staff removal, opening-based filled heads with stem checks, ring-tested hollow heads, stroke-structure accidentals with re-pairing of fragments cut by staff removal, position-checked key signatures with page consensus, system barlines). Ground truth from Verovio engravings in 5 fonts through a deterministic photo simulator: 99.7% heads / 99.2% pitch clean, 95.5% pitch on simulated photos, 89% on low-res phone shots, all staves and clefs found; ~300 ms a page. Sanity-checked on a real 1920s Haydn quartet scan (16 staves, all clefs). bun test (18), node smoke.js (17 checks through the real UI, export and offline PWA).

30 Sep 2026

StaffInk slurs + articulations: accent, tenuto, marcato, staccatissimo and fermata recognised from shape (dash, tick, wedge, caret, arch + dot) and position (just above/below one note, clear of an up-stem), which separates an accent from a hairpin and a tenuto from a ledger line; a mark written straight after its note splits off the pending note first. Engraved on the notehead side, stacked outward (Gould), accents kept off the staff lines, fermata above the staff and any slur. Slurs: notehead side / above for mixed stems, flip button, arch raised to clear inner notes, drawn in pieces across line breaks. Scribble or eraser removes a note, else just its mark, else a slur. Playback: legato under slurs, small gap otherwise, tenuto full, staccatissimo 25%, accent/marcato louder, fermata doubles its note and shifts everything after. MusicXML articulations/fermata/slur placement. 54 tests, 27 smoke checks.

29 Sep 2026

Add StaffInk - stylus/touch music notation writer. Fact check: StaffPad (2015 Windows, 2020 iPad, Apple Design Award 2020, Muse Group since 2021) is the leading handwriting-first notation app, but NotateMe (Neuratron, 2013, iOS + Android) came first, and the notation market overall belongs to Sibelius, Dorico and MuseScore (Finale was sunset Aug 2024). Surveyed Presto (OZCHI 1996), the Music Notepad (UIST 1998), Miyao & Maruyama’s stroke-primitive recognizers, the HOMUS dataset (ICPR 2014) and the $P/$Q recognizers.
StaffInk recognizer: $P nearest neighbour on HOMUS scored 64% writer-independent; a 171→128→64→32 MLP over directional element features + staff-relative size, trained in plain JS with augmentation (train.js, zero deps), scores 94.4% top-1 / 99.3% top-3 on held-out writers 81-100. Notes are then parsed structurally (stem, heads at one end, fill density) for pitch and chords; beams, ties/slurs and scribble-erase are gestures checked against what is already engraved, with pending ink committed tentatively first. Alternatives strip after every symbol; a tap fixes it and stores a $P template so recognition adapts.
StaffInk tuplets, dynamics, hairpins: write a 3/6 over notes (window of consecutive notes whose lengths add up to n units; number on the beam side, bracket unless it is one beam, own beam group), pp–ff/sfz/fp under a note, </> in one stroke or two lines. Tuplet digits are $P templates cut from real HOMUS time-signature numerators (train.js --digits); dynamics are synthesised italic-letter templates (weakest part, corrections take over). Only matched outside the staff or beyond the stems, with a guard for whole notes on ledgers. Dynamics and hairpins drive playback velocity; MusicXML time-modification/tuplet/dynamics/wedge. Edit bar gets a triplet button and a dynamics row.
StaffInk offline PWA: manifest + icons, service worker precaching all 20 files (stale-while-revalidate). Smoke test reloads with the network emulated offline and writes a note. 44 tests, 23 smoke checks.
StaffInk engraving: SMuFL anchors and engraving defaults, logarithmic spacing, bars reserve room for missing beats, greedy line breaking + justification, beaming by beat, seconds in chords, stacked accidentals, accidental carry and ties in pitch spelling, grand staff alignment by onset. Web Audio playback, MusicXML 4.0 / MIDI / PNG export. bun test (37) + node smoke.js (18 checks with real pen and touch pointer events over CDP).

26 Sep 2026

Add Marine AR - AR vessel finder built on the estate-ar pattern. Surveyed live AIS sources reachable from a browser: Digitraffic (Finland) is the only public CORS REST feed; aisstream.io is worldwide but bans browser connections so proxy.js re-serves it in Digitraffic’s shape; US live AIS (NAIS / MarineCadastre) is not public. Lanes come from NOAA ENC Direct TSSLPT/TSEZNE/PRCARE/TSSBND/FAIRWY layers (harbour + coastal, verified CORS). Pinhole camera with roll, near-plane clipping so lanes you stand in still draw, curvature-dropped sea surface, true horizon dip, hull footprints from A/B/C/D, dead reckoning, CPA/TCPA. bun test with captured fixtures + headless Chromium smoke render.
Marine AR land occlusion: ENC Direct LNDARE polygons (server-generalised via maxAllowableOffset), 2D sight-line vs shoreline test per vessel at 2 Hz. Ships behind a headland render dimmed/dashed with distance to the blocking shore; lane labels behind land are dropped. Viewer standing on land is handled by skipping the first crossing of the ring that contains them.
Marine AR elevation-aware occlusion + AtoN: LNDELV spot heights attached to land rings, sight line from eye height vs land height vs estimated ship air draught with curvature; a lighthouse on the shore is no longer hidden by its own shore. Buoys, beacons, daymarks and lights from ENC Direct pinned in AR with S-52-ish glyphs and IHO light characters (Fl R 2.5s 5M), tap for details. Smoke test rewritten over the DevTools protocol (fixed settle instead of network idle, reads page state).
Marine AR live AIS everywhere, no key: OpenSeaFeed (community AIS network, aisstream.io protocol) over a keyless WebSocket straight from the page; browsers skip CORS for WebSockets so no proxy is needed. Probed the free tier at 30 000 deg² of bounding box; an AR view uses ~0.2. Auto now picks national feed → OpenSeaFeed → demo, the demo chip says SIMULATED, and with hundreds of real marina boats behind the shore only the six nearest occluded ships get labels.
Marine AR: vessel chip now splits the count into underway and “N moored hidden” (tooltip explains the show-moored setting). Live Seattle names/types/speeds cross-checked against VesselFinder detail pages: HUDSON (tug, 7.5 kn to Tacoma), TANGO, MAVERICK, PALADIN, BADFISH, BARONESS all agree.
Marine AR: vessels whose static report has not arrived read “Unnamed tug” / “Unnamed vessel · US” (flag from the MMSI prefix) instead of a bare MMSI; vessel card gains a “VF map” link that centres VesselFinder’s live map on the ship.
Marine AR night lights: after dark (sun < −3°, computed on-device) every charted light flashes its real character from S-57 LITCHR/SIGPER/SIGGRP (Fl R 2.5s, Fl(2+1) 15s, Oc, Iso, Al WR, Q+LFl…) with a range-sized glow and its own phase. Silent vessels get their names from OpenSeaFeed’s per-vessel endpoint (250 B each, cached 30 days) instead of a 5 MB snapshot. Scanned from Alki Point: WSF CHIMACUM/KITTITAS/WENATCHEE, SALLY FOX, ENETAI on the ferry routes, all live.
Marine AR fix: AR view had dropped to ~2 fps so labels barely followed the phone. Lane polygons crossing the near plane projected to ~10 million px and their dashed outlines expanded to ~46 million px of dashes per frame; projected rings and lines are now clipped to a padded screen rect (60 fps restored). Also: relative-only orientation (Firefox Android) no longer freezes after the first event, a compass that wakes after start-up leaves drag mode, pose smoothing is time-based, and moored boats’ line of sight refreshes every 10 s instead of twice a second.
Marine AR fix (iOS Safari): labels stayed put when moving the phone because motion access was requested after awaiting the camera, by which time Safari no longer counted it as part of the Start tap and refused, leaving drag mode. The request now happens synchronously in the tap; if it is still refused the HDG chip reads “HDG off · tap to allow” and tapping it asks again.
Marine AR: Detail slider (Minimal → Everything) to declutter, label slots prioritised to big and moving ships; tap the radar to flip between AR and a full-screen 2D chart (land, lanes, aids, vessels/hulls, range rings, pinch-zoom, pan, north/heading-up).

26 Sep 2026

Add Estate AR - property price AR viewer. Surveyed sources callable from a static page: King County’s KingCo_PropertyInfo ArcGIS MapServer answers point+distance queries with WGS84 polygons, appraised land/improvement values and a 3-year sales layer, all with CORS; Nominatim for place names. geo.js does centroid/angular-span/point-in-polygon and a pinhole projection from (bearing, pitch); Android Euler angles are rotated to the back-camera direction so any phone orientation works. Labels, ground-plane outlines when tilted down, radar, detail sheet with sales history and assessor links, simulated mode for desktop. bun test (33) on a captured live fixture + live smoke test.
Estate AR: Snohomish County provider (pds_prop_report MapServer, market land/impr values) and the WA State Parcels Project layer on DNR’s server as the fallback for Pierce and every other WA county (values match King County’s own service exactly; DOR land-use codes decoded). Labels moved to rooftops: building height from OSM height/building:levels via Overpass (CORS works from a browser; 37/134 buildings mapped in the U-District), else estimated from use class + improvement value per ft² of lot (log fit through $60→1, $1000→6, $8500→24 storeys; reproduces the Brooklyn Ave skyline). Dashed roof outlines + corner verticals. PWA: manifest, icons, service worker (shell cache-first, data network-first with fallback), IndexedDB parcel store with fetched-coverage circles so revisited streets work with no signal. Verified offline reload in headless Chromium. 48 tests.
Estate AR: radar and a settings-sheet mini-map show the offline coverage union (wash + exact boundary from unionArcs, which computes the uncovered arc intervals per circle — canvas destination-out tricks left antialiasing ghosts) and amber dotted circles for pending downloads; “Precache around me” with a slider, request estimate and a hex-lattice planner (planPrecache, tested gap-free) that skips already-covered circles and runs two requests at a time; 0.5 mi of the U-District = 19 requests, ~3k parcels, 5 MB, ~2 s. Service worker shell strategy changed to stale-while-revalidate so edits reach installed copies. 53 tests.
Estate AR: coverage wash drawn on the ground plane in the camera view (circles clipped to the view wedge with Sutherland–Hodgman before projecting, near-plane vertices pinned to the screen bottom, union boundary stroked as a line on the pavement); auto-precache ahead along the GPS course (Track needs 30 m over 90 s, planAhead lays a 3-lane corridor of fetch circles, allowAuto gates on navigator.connection type / saveData with Off / Wi-Fi / Any), sharing one download runner with the manual button. 59 tests.
Estate AR: optimisation pass. Bandwidth: server-side generalisation (maxAllowableOffset 0.5 m, 52→18 KB gzipped per circle), paging in dense cores, Overpass skipped when heights are already known, rolling 24 h data meter with a daily budget gating auto-precache. CPU: rings converted once to a metric frame and decimated, planar parcelViewXY replaces per-vertex haversine, 30 fps cap with change detection, radar footprints on a cached offscreen bitmap, union arcs cached — main thread 0% idle → 96% idle still / 74% turning with 740 parcels. Battery: 720p camera, sensors and camera suspended when hidden, low-battery mode via Battery API. Large commercial: stacked condo/air-rights parcels merged into one label with summed value and a units list, wide parcels anchor the label where you are looking, county-server outage falls back to the statewide layer (Snohomish went down mid-session). 68 tests.
Estate AR: per-unit economics for large buildings. King County’s nightly Apartment Complex + Unit Breakdown extracts (no CORS, 9,170 buildings) and the WA rows of Zillow Research’s ZORI/ZHVI ZIP CSVs are folded by fetch_data.js into two small bundled JSONs; market.js gives assessed value per unit, $/ft² living, unit mix, year built, ZIP typical rent/condo/home with 1y change, and an estimated rent per unit (ZIP rent × size^0.6 × age factor) with gross yield. Assessor storey counts now override the height estimate. “Per unit” price mode. The detail sheet states what assessed value is (1 Jan assessment-year estimate, not last sale). 76 tests.
Estate AR: Los Angeles County provider (Calabasas beta tester). Parcels from the county’s hosted layer with roll values, Prop 13 base years, living ft², beds/baths, units and year built; recorded sales from the assessor’s PAIS layer. Because Prop 13 pins assessed value to the purchase price, California defaults to a comps-based market estimate (median $/ft² of nearby 4-year sales × living area, with IQR) and the sheet explains the base year. Zillow ZIP indices now bundled per state (WA, CA); rent estimate for single homes; ?sim=calabasas. 81 tests.
Estate AR: elevation. terrain.js looks up ground elevation per parcel (Open-Meteo, 100/call) and under the viewer (USGS 3DEP), cached on a 30 m grid, so hillside lots sit at their real height; GPS altitude is geoid-corrected, ground-subtracted and fused across fixes weighted by altitudeAccuracy, trusted only under 12 m uncertainty (“~floor 25 ±1” in the chip), with a manual Floor override in settings. From a high-rise the street grid now drops below the horizon instead of every label sitting 30° too high. ?sim=…&floor=24 / &alt=166. 87 tests.
Add SailNav - studied OpenCPN (s52cnsy DEPARE01 banding, georef Mercator sailing, Routeman XTE/arrival) and built a browser plotter on NOAA ENC Online tiles + ENC Direct DEPARE/SOUNDG queries. Depth-under-boat with tide correction, grid A* water routing respecting draft, shore margin and charted hazards. bun test + live NOAA smoke test.
SailNav offline: s57.js reads ISO 8211 / S-57 .000 cells directly in the browser (DecompressionStream unzip, DataView parsing, IndexedDB store); vector S-52 chart, depth and routing without network. Verified identical to ENC Direct on US5SEAGK. Bundled 62 central Puget Sound cells (4.9 MB) in vibes/sailnav/charts/ with auto-load around the boat; fetch_charts.js refreshes them. Vendored Leaflet + service worker + manifest: whole app runs offline after first load (verified with network disabled in headless Chromium).

26 Sep 2026

Add West Point Buoy - single-file dashboard replacing the NDBC station page: NDBC 10-min wind/gust via tiny CORS proxy (NWS hourly fallback), NWS hourly forecast dashed onto the wind chart, pressure/temp, CO-OPS tide, parsed coastal waters forecast with advisories, wind rose histogram for the selected range (click a sector to filter the chart by direction), diurnal hour×day heatmap, side-by-side station compare overlay
Add tide predictor - harmonic tide prediction in plain JS. tides.js is a port of the Schureman SP-98 constituent definitions (following pytides and XTide’s congen_input.txt), fed by NOAA CO-OPS harmonic constituents. Visualizations: height curve with daylight bands and NOAA overlay, tide clock, hi/lo table, ±45 day spring/neap envelope with moon phases, constituent bars, decomposition of the top constituents, animated phasor sum.
test_tides.js validates against NOAA official predictions for 8 stations: height rmse 0.006-0.11 ft, hi/lo timing mean error under 1 minute. Gotcha found: NOAA’s M1 uses Schureman’s second formula (V = T - s + h + 90, u = xi - nu + Q) with the speed of the first (14.4966939 deg/h).
Research notes: surveyed pytides, UTide, pyTMD, PyFES/aviso-fes, hatyan, slackwater (ex-neaps, TS) and its station database, solunar. Picked pytides/congen as the reference for a small browser port; slackwater is the best maintained JS option if a dependency is acceptable.

17 Feb 2026

Add BrainWave - Neural Audio Engine - Web Audio app for brainwave entrainment with binaural beats, isochronic tones, and 9 study/focus mode presets
Procedural ambient music engine with 10 generative styles (ambient pads, piano, space drone, singing bowls, lo-fi, arpeggios, dark pad, cinematic lo-fi, cinematic arpeggio, shimmer)
Independent volume controls for master, beats, and music layers
Real-time waveform/frequency visualizer, session timer, 6 ambient noise generators

13 Oct 2025

Add CPU/Numa Affinity planner / visualizer

15 Jul 2024

Attempted to reproduce moon fraction vis using astronomy lib

14 Jul 2024

Testing astronomy engine lib. Also nodejs comparison with suncalc3.

11 Jul 2024

moon charts - strip down to basic 30 day dump.

10 Jul 2024

Visualization of moon across it’s phases. moonfraction. Other WIPs moonpath wip1 moonpath wip2

16 March 2024

Adding TextGeometry to three.js, including serialization/deserialization support https://github.com/mrdoob/three.js/pull/27931

11 Feb 2024

Terminal Mirror - Mirrors a tmux session to a web interface over fetch streaming. This was interesting as an “external monitor” in passthrough mode on my quest.

tmux mirroring pic

6 Feb 2024

Canvas version of the Starlings murmurations inspired by https://twitter.com/JulianGarnier/status/1754495444896416025 (incomplete, need to follow up on tweens)

5 Feb 2024

Simple visualization to work out earth, moon, and sun relations

2 Feb 2024

Added true north for compass heading, updated compass sun position to include current angle
Integrate “AR passthrough” for sun path ar 2d, removed WebGL renderer

1 Feb 2024

Added color palette test utilizing rgb, oklab and spectral interpolation

30 Jan 2024

sun path ar 2d alternative implementation of sunpath ar but rendered on 2d canvas. this uses threejs for 3d calculations but not rendering.

28 Jan 2024

updated Sun path viz with equal altitude spacing
Initial implementation of AR sunpath using threejs, geolocation and device motion
sun path ar for mobile and sun path three for desktop experimentations
sun year path - Sun’s visualization path of the year 27 Jan 2024
exploration of AR with compass direction. threejs device orientation test 26 Jan 2024
Compass with position, compass variant that uses geolocation for coordinates and altitude. Also combined sun positioning to show compass with sun position 24 Jan 2024
JS based generation and visualizations of sunrise and sunset tables. Explore the use of the suncalc libaries and geolocation api. 23 Jan 2024
A JS canvas based implementation of the IOS Compass app, utilizing device orientation event and webkit compass heading. I implemented most of the UI/UX including bearing display and the dial rotation, including bubble level and direction marking. What’s missing are vibration (disallowed in safari) and GPS coordinates and some subtle animations. One improvement I have over Apple’s implementation is that tapping the screen not only marks the change in direction, but gives the angle difference. This can be useful for just measuring differences in compass bearings. 14 Jan 2024
combined streaming server and tracking code in handstand experiment to mirror pose tracking across the network 14 Jan 2024
streaming-server uses h2 and sse (server sent events) as an alternative to websockets.