starlight

For anyone making changes (human or AI):

• Bump the version in package.json with every change.
• Add an entry to the ## Changelog section at the bottom of this file describing what changed and why.
• Saved mixes store only parameter state (never audio — loops re-render fresh every session), and storage/migrations.ts drops unknown sound ids and unknown-shaped data safely. So retuning a sound or renaming a param never needs a storage change. Bump the saved-mixes key in src/storage/keys.ts (STORAGE_KEYS.savedMixes, e.g. -v2 → -v3) only if a kept param key changes meaning incompatibly (same key, different effect) — and know that bumping discards users' saved mixes, so prefer a migration.
• If you add or change an icon glyph, follow Icons: the Material Symbols subset below — new glyph names must be added to the subset font or they render as text.
• If you add/remove a sound or change a headline feature, keep the discoverability surfaces in sync: the meta/OG descriptions and the JSON-LD featureList in index.html, and public/llms.txt (see Discoverability below).
• UI changes must pass the accessibility gate: tests/e2e/a11y.spec.ts fails on any serious/critical axe-core violation across the primary surfaces.

starlight (repo SleepSounds; formerly "drift away") is a mobile-first ambient sound app for relaxation and sleep. All 19 sounds are generated in the browser; nothing is streamed or downloaded. Rain, Thunder, Windy Forest, Fire, and Birdsong are synthesised live via AudioWorklet (event-based); the rest are procedural WAV loops. (Two more finished sounds, Stream and Shower, are pulled from the lineup for now — HIDDEN_SOUND_IDS in src/data.ts — and don't count toward the public tally.)

Features

• Scenes: ten curated mixes as gradient-art cards; tap to play instantly.
• The library: 19 procedurally generated sounds across eight categories, layered freely with per-sound volume. All but the noise trio open an editor that leads with named character presets (variant chips with small drawn marks — e.g. ocean's Lapping Shore / Distant Surf / Rolling Waves / Storm Surf) and hides fine-tune sliders behind a toggle.
• Mini player + now-playing sheet: persistent player bar; the sheet holds per-layer sliders, master volume, sleep timer, save-mix, and drift mode.
• Drift mode: fullscreen night surface with clock, breathing play orb, and screen wake lock for the nightstand.
• Sleep timer (15m/30m/1h/2h/4h/8h) with a progress ring and a gentle 90-second wind-down fade; the night sky dims with it.
• Your mixes: save/load/delete via localStorage; lock-screen media controls (Media Session API); installable PWA with custom install prompt.

Tech Stack

• React + TypeScript + Vite
• Hand-rolled CSS design system (src/index.css, documented in DESIGN.md); Tailwind present only for its base reset
• HTMLAudioElement + Web Audio worklets for playback and mixing
• Code-split for a light first paint: the WAV-generator DSP, the now-playing sheet, drift mode, and the sound editor are fetched on demand (and idle- prefetched after first paint), so the initial bundle carries only the shell.

Project Structure

See ARCHITECTURE.md for how these fit together (audio graph, interruption handling, storage, PWA update flow, testing).

starlight/
├── index.html
├── package.json
├── postcss.config.js
├── tailwind.config.ts
├── tsconfig.app.json
├── tsconfig.json
├── tsconfig.node.json
├── vite.config.ts
├── docs/ADDING-A-SOUND.md  # the checklist for adding a generated sound
├── docs/research/          # background research reports (reference only)
├── tests/                  # vitest unit tests + tests/e2e Playwright specs
└── src/
    ├── App.tsx              # shell: scenes, mixes, library, player state
    ├── data.ts             # sound library + built-in presets (lazy WAV loader)
    ├── index.css           # the design system (see DESIGN.md)
    ├── main.tsx
    ├── types.ts
    ├── audio/
    │   ├── dsp.ts               # noise sources, filters, loop-conditioning, WAV encoding
    │   ├── generators.ts        # per-sound procedural WAV synthesis (code-split)
    │   ├── graph.ts             # the shared AudioContext + master bus + interruption guard
    │   ├── layerMeta.ts         # masking roles + per-layer shaping (anti-fog)
    │   └── sources.ts           # mixer sources: crossfade WAV + worklet w/ fallback
    ├── components/
    │   ├── CookieNotice.tsx     # one-time storage notice
    │   ├── DriftMode.tsx        # fullscreen night surface (lazy)
    │   ├── ErrorBoundary.tsx    # last-resort crash guard
    │   ├── InstallPrompt.tsx    # PWA install affordance
    │   ├── MiniPlayer.tsx       # persistent bottom player
    │   ├── MixControls.tsx      # per-layer mixer (volume, mute/solo, save)
    │   ├── NightSky.tsx         # living canvas starfield
    │   ├── NowPlayingSheet.tsx  # mix control room (lazy)
    │   ├── SidePanel.tsx        # desktop docked mixer
    │   ├── SoundCard.tsx        # library tile
    │   ├── SoundEditor.tsx      # per-sound parameter editor (lazy)
    │   ├── Toast.tsx            # forgiving undo snackbar
    │   └── soundEditorDefs.ts   # editor models: groups, variants, defaults
    ├── hooks/
    │   ├── useAudioMixer.ts     # playback engine wrapper (state ↔ audio sources)
    │   └── useSleepTimer.ts     # sleep-timer countdown, wind-down fade, handlers
    ├── storage/                 # localStorage keys + load/save + migrations
    ├── platform/                # platform seam (web bridge today)
    ├── config/                  # feature flags
    ├── utils/                   # logger
    └── lib/
        ├── backgroundAudio.ts · categoryIcons.ts · haptics.ts · scenes.ts
        ├── sliderFill.ts · soundIcons.ts · time.ts · variantIcons.tsx
└── public/worklets/             # live AudioWorklet generators (rain, fire, …)

Setup

1) Install dependencies

npm install

2) Start development server

npm run dev

3) Build for production

npm run build

4) Preview production build

npm run preview

Developer tips

• Dev mode: five quick taps on the moon (within 3 seconds) toggle it — the moon wanes to a crescent, the greeting says so, and the library reveals the hidden / pulled sounds (e.g. Shower, Stream) for auditioning. Five more taps or a refresh turns it off; nothing persists.
• Tests: npm run test (vitest units), npm run test:e2e (Playwright over the production build — smoke flows plus the axe-core accessibility gate), npm run lint.

Icons: the Material Symbols subset

The UI's icons are Google's Material Symbols Rounded, self-hosted as a ~7 KB subset (public/fonts/material-8.woff2) of the ~510 KB full font, so icons work offline and no request leaves the device. It's a ligature font: markup contains the icon's name as text (<span class="material-symbols-rounded">pets</span>) and the font maps that letter sequence to the glyph.

Whenever you reference a new glyph name anywhere in src/ — the icon maps in src/lib/soundIcons.ts / src/lib/categoryIcons.ts, or an inline material-symbols-rounded span — you must add it to the ICONS list in scripts/subset-icons.py and re-run the subset:

pip install fonttools brotli
python3 scripts/subset-icons.py

A glyph missing from the subset renders as its raw name in text (the word "pets", not tofu), so eyeball any new icon. The script fails loudly on a typo'd name. In-place font updates are cache-safe: the service-worker build id hashes every precached file's bytes (scripts/inject-precache.mjs) and the SW installs with cache: 'reload' (public/sw.js) — see the script's docstring for the history behind that.

Discoverability: SEO, social cards, and AI readers

The app is a client-side React shell, so scrapers that don't run JavaScript see nothing in the body. Everything they need lives in explicit surfaces, which must be kept in sync by hand when features change:

• index.html — the meta description, the Open Graph / Twitter card tags (absolute URLs; WhatsApp requires them), and one JSON-LD schema.org/WebApplication block (keep it to exactly one entity).
• public/og-starlight.jpg — the 1200×630 social preview, generated from the app's own ingredients (starfield, moon, Cormorant wordmark) by scripts/og-card/render-og-card.mjs; re-run it rather than editing the JPEG.
• public/robots.txt + public/sitemap.xml — everyone welcome, AI crawlers included; the sitemap lists the app and the privacy page.
• public/llms.txt — a plain-text summary for AI agents (per the llms.txt convention): what the app is, the feature list, key facts, doc links.

Notes

• All audio loops are procedurally synthesized in src/audio/ (generators.ts over the dsp.ts helpers) and encoded to WAV blobs at runtime (two-channel 16-bit PCM; a few non-directional sources stay centred). Stereo width is baked in at generation — broad beds via decorrelated opposite time-shifts (no comb filtering), discrete events via equal-power panning — and layering is masking-aware (layerMeta.ts roles + layerShaping). The generator module is code-split, so it loads on a sound's first play rather than at startup.
• No backend is required.
• The version number (from package.json) renders inline in the page footer (.footer-meta in src/App.tsx), beside the privacy link.

Changelog

0.1.6

• Fix: no more opaque/normal background flashing on drop. The lifted card deliberately carries an opaque surface (so cards it passes over don't show through it), but across the drop the drag classes were snapping that background opaque → normal → opaque → normal, each switch instant — the "flashing" reported on device. The shade story is now continuous: the card fades to opaque on lift (background joined the lift transition), stays opaque through the entire commit swap (the lifted look is held across the atomic reorder), and then makes one smooth fade back to the normal translucent surface during the landing glide — the landing class is no longer opaque and transitions background/border along with its transform. Verified under 6× CPU throttling by sampling the card's computed background alpha every frame after release: a single monotonic fade (1.00 → 0.06), never rising back toward opaque. Gate: 118 unit, 30 e2e.

0.1.5

• Rearchitected the drop: commit at release. 0.1.4's atomic swap wasn't enough on device, because the architecture still had a swap moment — a 190 ms landing glide over the old DOM order, then a handoff. Anything that slipped in that window (a paint, a stretched frame, cached-rect drift) moved cards that were already in place. The drop is now structurally incapable of that: the instant the finger lifts, the new order commits synchronously (flushSync) and every transform comes off in the same paint — the displaced cards' new layout positions are by construction exactly where their shift transforms already had them, so that swap is invisible — and then the dropped card alone FLIPs from its position under your finger to its real new layout position (measured from the committed DOM, not cached rects) in one glide. No timers before the commit, no geometry math at the end, nothing left to race. The interim drag-drop snap mechanism from 0.1.3 became unnecessary and was removed. Verified under 6× CPU throttling in two scenarios (instant release; release while neighbours were still mid-glide): displaced cards show zero movement from the moment of release, the dropped card makes exactly one glide, and nothing moves after it lands. Gate: 118 unit, 30 e2e.

0.1.4

• Fix: nothing moves after the drop lands — the swap is now atomic. Follow-up to 0.1.3 from on-device testing: cards that were already visually in place at release still slid/jumped into position again a moment later — both the dropped card and the displaced ones. Root cause: the final handoff wasn't atomic. After the landing glide, the code cleared every inline transform immediately (snapping all cards back to the old DOM layout) but asked React to commit the new order asynchronously — on a slower device a paint slips into that gap, so the whole grid flashed back to the old arrangement and then moved into place again. (Desktop committed fast enough to hide it, which is why earlier traces looked clean.) The commit now runs synchronously (flushSync) with the transform-clear in the same JS turn under the settle guard, so the browser paints exactly once across the swap and the visual layout never changes. Verified under 6× CPU throttling (phone-like commit latency): zero movement on any card after the landing completes, correct final order. Tagged [0.1.4] in useGridReorder.ts.

0.1.3

• Fix: no more side-wipes after a drop, and no wrong-place wipes. Two more drag-feel bugs from feedback:
  1. Displaced cards finished their slide after the finger lifted. The neighbour-shift glide (180 ms) triggered by the final pointer movement was still running at release, so "after a move the displaced card does a side wipe into place". Now, the moment the finger lifts, every displaced card snaps to its final cell (a drag-drop class suppresses their transition) — only the dropped card itself glides in. Measured: zero post-release movement on any non-dragged card, where previously the wipe ran ~60–180 ms past release.
  2. Dragging with a sound editor open wiped cards to positions from the wrong layout. The hook cached every card's rect at lift and then asked the app to close the open inline editor — which relayouts the whole grid, leaving every cached rect stale by the editor's height. All shift and landing glides then animated to coordinates from a layout that no longer existed (reproduced: cards gliding to the top of the screen). The close now happens synchronously before measurement (flushSync in the lift callback, measurement after), so the geometry is always the real one. Both cases are covered by the trace-based verification and a new e2e guard.

0.1.2

• Fix: drops land on the box the card visibly covers. Feedback with a screenshot: drop targeting "favours the right too much — even more so when it's the right one on a lower line". Three compounding causes found and fixed in the drag hit-testing:
  1. The row-major probe ignored x below a row's centre-line. The old score (y·K + x) meant hovering the bottom half of any card counted its entire row as "passed" — the target jumped a whole row regardless of which column you were over. That was the "worse on a lower line" bias.
  2. The probe was the finger, not the card. Grab a card by its corner and the finger sits half a card away from what the card visibly covers, so drops skewed toward the grab side.
  3. Auto-scroll drifted the target under a stationary finger. The old linear ramp (68 px zone, up to ~12 px/frame immediately) meant dropping on a card near the bottom of a small phone screen scrolled the list away mid-drop (reproduced on the Pixel-profile viewport: the slot ran from 2 to 7 while simply holding). Now the card's visual centre claims the cell it covers, Voronoi-style — rows banded first (boundaries midway between row centre-lines), then the nearest column, so the decision boundaries sit in the gaps between cards, exactly where the eye expects — and auto-scroll ramps quadratically from zero at a smaller (44 px) zone edge with fractional accumulation, so a drop near the screen edge barely drifts while a push to the very edge still travels fast. Verified on the two reported scenarios (corner-grab onto a lower-left cell; bottom-half hover on a same-row right cell) plus a new permanent e2e test on the real device profile.

0.1.1

• Fix: dropping a dragged card no longer flashes cards "in" as if new. Listening-to-the-UI feedback on 0.1.0: right after letting go, cards near the drop slot flashed like fresh arrivals even though nothing was moving. Mechanism: the entrance reveal (fadeSlideUp with backwards fill and a per-card stagger delay) is a finished CSS animation — and a finished CSS animation restarts whenever its DOM node is moved (React moves the dragged card's node when the reorder commits) or its animation declaration is re-resolved (the drag-lift class toggles animation: none on and off). A restart under backwards fill means the card first blinks invisible for its whole stagger delay, then fades in from below — "as if new". The reveal is now frozen with an inline animation: none the moment it finishes (onAnimationEnd, self-target guarded), making it one-shot for the life of the node while fresh mounts (category switches) still animate. Verified in the real browser: before, the dropped card sat at computed opacity 0 ~300 ms after release; after, zero cards are below full opacity on short and long drags, all 19 animationend freezes fire, and the initial reveal still plays. Guarded by a new e2e test.

0.1.0

• Arrange the library: drag-and-drop sound reordering. Hold any sound card (~⅓ s, touch or mouse) and it lifts — opaque, raised, following your finger — while its neighbours glide aside to show where it will land; release to drop, Escape (or a cancelled gesture) glides everything home. The arrangement is yours: it persists on-device (drift-sound-order, loaded through the same never-throws migration contract as the rest of storage — unknown ids are dropped, sounds added in later releases append rather than vanish) and works the same inside any category filter (a move within a filtered view splices correctly into the full order). The details that make it complete rather than a proof of concept:
  • Scroll-safe: moving before the hold completes cancels the lift, so scrolling and tapping are untouched; while dragging, native touch scroll is suppressed and dragging near the top/bottom edge auto-scrolls the page.
  • No stray toggles: the release click is swallowed, so a drag never turns the sound on; lifting closes an open sound editor; the long-press context menu and text selection are suppressed.
  • Accessible: every card carries a keyboard grip (hidden until keyboard-focused) — arrows step the card through the grid, Home/End jump — and every pick-up, move, and drop is announced to the live region. Axe scans stay clean.
  • Calm motion: slot-shift and landing glides are short eased transforms over rects cached at lift (no mid-drag layout thrash), the commit swaps transformed-old-DOM for clean-new-DOM with transitions suppressed for one frame (no flicker), and prefers-reduced-motion gets instant snaps.
  • Discoverable: the library section header now reads "19 sounds · hold to arrange". Covered by new unit tests (order storage migration, move semantics) and two e2e tests (a real pointer drag with reload persistence + the keyboard grip path); llms.txt and the JSON-LD feature list mention the capability.

0.0.47

• Ticking Clock: less wood, more mechanism. Listening feedback: the tick "sounds too much like wood" — and acoustically that was exactly right: a single exponentially-decaying sine is the woodblock model. Metal reads as inharmonic partials, so each tick now rings two faint, faster-decaying partials at 1.63× and 2.47× the beat's own frequency (per-beat micro-detuned ±1 % with random phase, so no two ticks shimmer identically), under the unchanged fundamental and contact noise. The partials sit below the top-safety shelf, so the voicing stays gentle. Beat structure is untouched (tick counts byte-identical per variant: 19/34/34/60) and per-variant RMS is within 0.9 dB of 0.0.46. The pace / tick–tock / brightness sliders and the bare-mechanism character are exactly as before — the material of the mechanism just reads as brass, not birch.

0.0.46

• Wind Chimes: no longer opens on the end of a ring. Listening feedback: the chimes "start at the end of a ring every time". Two causes stacked: the loop blend crossfades the buffer's end into its first 1.2 s (so the start carries a dying ring tail you never heard struck), and the gust²-gated scheduler could leave the first real strike 4–22 s away (measured per variant). Three changes: (1) strike chance and strength now taper away over the ~3 s before the seam guard, so the material the blend copies forward is near-silence (seam-tail RMS 4–8× quieter per variant); forced lull-bridging strikes skip the taper zone. (2) A guaranteed opening strike (with a likely second swing, so it doesn't read staged) lands at ~1.7–2.6 s — placed deliberately past both the mixer's 1.6 s fade-in (so it's actually heard on first play) and the 1.2 s blend zone (so it isn't copied into the seam). (3) The ghost residue before that first strike drops −8…−12 dB. Measured: first onset 4.3/4.8/4.8/22.4 s → 1.9/2.2/2.0/2.2 s across the four variants; RMS within ±0.9 dB; loop-seam continuity untouched (the blend still runs — there's simply nothing left for it to smear).

0.0.45

• Train: the keyboard click is gone. Listening feedback: an occasional sharp click that "doesn't sound like anything on a train". Diagnosis: the joint-clack noise bursts started at full amplitude on their very first sample — an instant broadband edge whose spectral splatter cuts far above the clack's own lowpass and reads as a keyboard key, not steel on steel. It was the one discrete event left in the library violating the no-clicks rule (an eased onset on every event). Each clack now opens with a ~1.2 ms raised-cosine attack. Measured: the worst-case high-band edge (max sample step above 3 kHz) halves — 0.134 → 0.067 on the default, 0.166 → 0.074 on Old Local — while clack-band and overall RMS stay within 0.15 dB, so the rhythm and level are untouched; only the false edge is gone.

0.0.44

• Ticking Clock: the second contact rings too. The escapement's secondary contact (~11 ms behind each beat) was rendered as a pure noise burst — a "tk…sh" where a real mechanism goes "tk…t": the second contact excites the same plate/gear resonance as the main one, only weaker and darker. The echo now carries a small damped ring at 0.85× the beat's own frequency under a reduced noise share (energy split so its total level is unchanged), and its delay wobbles ±1.5 ms per beat (mechanical variation — no two beats land identically). Beat structure is untouched (tick counts byte-identical per variant: 19/34/34/60) and per-variant RMS within 0.6 dB. This closes the six-sound refinement pass (0.0.39–0.0.44); as its capstone, the live sweep (scripts/sweep-sounds.mjs) confirmed 19/19 sounds audible through the real master bus with no console errors. Research notes: docs/research/tail-sounds-refinement.md.

0.0.43

• Wind Chimes: strike points and clapper bounces. Two things a real clapper does that the render didn't. (1) It lands somewhere different every time: which transverse modes a hit excites depends on the contact point, so the mode balance is now drawn per strike — barely on the fundamental, progressively wider on the higher modes (±15 % → ±40 %), mean equal to the old fixed balance so the tone voicing is unchanged in expectation. Every note keeps its pitch; its sparkle varies naturally. (2) It bounces: a clapper swinging into a tube often re-contacts it 40–90 ms later at a fraction of the energy — the soft "to-tok" double every real set makes. 22 % of strikes now carry one (never inside the loop-seam guard). Systematic energy delta ≈ +0.3 dB (jitter variance + bounce energy); measured variant levels stay inside the family's natural render-to-render spread, and the event-scheduling, tuning, decay physics and stereo hang are untouched. Research notes: docs/research/tail-sounds-refinement.md.

0.0.42

• Cat Purr: a body that shifts. The purr's pitch, level and exhale share already varied per breath, but the body didn't: the chest/throat resonance the glottal pulses ring was one fixed frequency for the whole 32 s — a cat bolted in place. A real cat settled against you shifts a little with every breath, and the resonant colour moves with it. Each breath now draws its own resonance multiplier (±4 %) and throat-partial weight (0.32–0.44, mean equal to the old fixed 0.38), so the timbre breathes with the animal. Pulse rate, breath envelope, muffle voicing and the Sissom-calibrated ~24–26 Hz fundamental are untouched; per-variant RMS within 0.5 dB of 0.0.41 (the purr-to-heartbeat loudness calibration holds). Research notes: docs/research/tail-sounds-refinement.md.

0.0.41

• Heartbeat: respiratory sinus arrhythmia. A resting heart doesn't jitter white — its rate swings smoothly with breathing (faster on the inhale, slower on the exhale), a few percent either way over each ~4 s breath, and the effect is strongest in exactly the relaxed, drifting-off state this sound exists for. The old render drew each beat interval i.i.d. (±2.25 %), so successive beats were uncorrelated. The interval now rides a slow loop-closed wander (±4 %, breathing-paced holds) with only a small white tremor (±1 %) on top. Verified on the raw interval series (crisp-onset render): baseline wiggles white around 924 ms with no structure; the new series undulates in smooth multi-beat waves (~891 → ~954 → ~893 ms) — a body breathing, not a metronome with noise. Mean rate, S1/S2 voicing and timing untouched; per-variant RMS within 0.3 dB. Research notes: docs/research/tail-sounds-refinement.md.

0.0.40

• Airplane: the twin-engine beat. Two jets in cruise never hold exactly the same shaft speed, so their cabin tones sit a fraction of a Hz apart and beat — the slow wah…wah swell that is one of the most recognisable cabin signatures (it's what prop synchrophasers exist to tame). The render had a single fundamental plus a 2.02× partial — no physical beat, just a random amplitude LFO (measured 60–160 Hz AM index 0.02–0.04 with no stable rate). The fundamental is now two engines 0.3125 Hz apart — ten exact steps of the 1/32 Hz loop grid, so the beat phase closes over the 32 s seam — each at 1/√2 of the old amplitude (total tonal power unchanged). Measured after: AM index 0.10–0.12 with a stable detected rate of 0.30–0.33 Hz on every variant — a gentle ~3 s swell, present but far below throbbing. Overall levels within 0.5 dB of 0.0.39 on all variants. Research notes: docs/research/tail-sounds-refinement.md.

0.0.39

• Train: the real four-beat clickety-clack. On jointed track a passenger coach rides on two bogies (two axles each) whose centres sit ~15.7 m apart under a ~20 m body, on ~19 m rails — so each rail joint fires four clacks (leading pair, then the trailing pair), and because the bogie spacing is close to the rail length, the trailing pair of one joint lands just before the leading pair of the next: the classic da-da … da-da …… gallop. The old render fired a single axle pair per joint — a two-beat, which is the rhythm of a two-axle freight wagon, not a coach. Verified structurally: the inter-clack gap histogram now shows the 0.40–0.45 s bogie-spacing cluster, which cannot occur in the single-pair render. Level-matched: per-clack strength ×0.72, floor-thump chance 0.4 → 0.28, and encode gain 0.68 → 0.60 (the softer four-clack stacks lowered the raw crest, so peak-normalisation had run up to +2 dB hot) — every variant lands within ±1 dB of its 0.0.38 loudness with the clack:floor mix ratio preserved to three decimals. Research notes: docs/research/tail-sounds-refinement.md.

0.0.38

• Build toolchain: vite 5 → 8 (Rolldown), audit now fully clean. Upgraded vite 5.4 → 8.1 (the Rolldown-bundler line), @vitejs/plugin-react 4 → 6, and vitest to 4.1.10 — clearing the last npm audit findings (the esbuild dev-server advisory, GHSA-67mh-4wv8-2f99): 0 vulnerabilities. The build output is equivalent: same entry/lazy chunks at comparable-or-smaller sizes, plus one new tiny shared chunk (useFocusTrap, ~0.8 kB, split because both lazy modals import it) which the SW precache injector picked up automatically. Verified end-to-end on the new toolchain: typecheck, lint, 110 unit tests (including generator determinism — WAV output is byte-identical, no sound changed), all 24 e2e including offline/service-worker, plus a dev-server smoke test. Follow-ups included: CI and deploy workflows bumped Node 20 → 22 (vite 8 requires ≥20.19; Node 20 EOLs April 2026), and package.json gained an engines field making the requirement explicit.

0.0.37

• Dependency hygiene: patched the @babel/core advisory. npm audit fix (non-breaking) resolved GHSA-4x5r-pxfx-6jf8 (arbitrary file read via sourceMappingURL) in @babel/core, a build-time dependency of the React plugin. No app code changed; the version bump exists so the rebuilt bundle flows through the normal update path. The remaining audit findings are the esbuild dev-server advisory (GHSA-67mh-4wv8-2f99), which affects only vite dev on a developer machine — the shipped app is static files and never runs esbuild — and whose only fix is a breaking vite 5→8 migration, deliberately not taken here.

0.0.36

• Perf: the last always-mounted surfaces stop over-rendering. Completes the memoization pass (0.0.11 / 0.0.21 / 0.0.35): DriftMode stays mounted after its first open, so every App render re-ran it even while closed — it's now memo'd, with its mixNames prop stabilized via useMemo (the inline activeSounds.map(...) minted a fresh array each render, which would have defeated the memo). NightSky likewise re-rendered on every App render though its three props are primitives that rarely change — now memo'd, so the sky only re-renders when playing / intensity / dim actually change; its rAF loop reads live props through a ref and is untouched. A timer countdown still updates drift mode each second while open. No behavioural or visual change. Tagged [v0.0.36 perf] in App.tsx, DriftMode.tsx, NightSky.tsx.

0.0.35

• Perf: the now-playing surfaces stop over-rendering. The mini player and the mix's layer rows (in the sheet and the desktop side panel) re-rendered on every App render — so a single volume drag repainted the mini player and every active layer row, not just the one being dragged, and a sleep-timer tick did the same. The mini player is now memo'd and each layer row is a memo'd LayerRow, fed constant-identity handlers from App (master toggle, stop, save, the timer actions, open/close, drift — ref-backed where they otherwise change with soundState or the timer). Measured with 5 active layers and a 10-event drag on one slider: the mini player re-rendered 0 times (was every event) and only the dragged row repainted — 10 row renders instead of ~50. Same behaviour, verified by the full e2e suite; extends the memoization pattern from 0.0.11 / 0.0.21 to the player surfaces. Tagged [v0.0.35 perf].

0.0.34

• Now-playing layer rows show each sound's own icon. The mix's layer rows (in the mobile sheet and the desktop side panel) labelled every layer with its category icon, so Fan, Thunder, and Wind — all "Air" — all showed the same glyph, while the same sounds carry distinct icons on their library cards, in the sound editor, and in the scene previews. The rows now use the per-sound icon too, so each layer is individually recognisable and iconography is consistent across the app. Same Material Symbols style and placement — a clarity refinement, not a visual redesign. Tagged [v0.0.34] in MixControls.tsx.

0.0.33

• Perf: layer masking recomputes only on active-set changes, not every volume tick. The mixer effect that applies per-layer masking (the gain-trim / low-pass / high-shelf that keeps stacked broadband soft) re-ran on every soundState change — including every frame of a volume-slider drag — even though the shaping depends only on which layers are playing and sleep-safe mode, never on their levels. Its own comment already said "recomputed whenever the active set changes"; now the code matches: shaping reads soundState through a ref and its effect is keyed on the joined enabled-id set (a string, so an unchanged set is a no-op dependency), while per-layer volume application stays on its own effect. The applied shaping targets are identical, so this is a pure work reduction — no audible change (verified by the audio-through-bus and sleep-safe e2e checks). Tagged [v0.0.33 perf] in useAudioMixer.ts.

0.0.32

• Cleanup: remove the dead playback-rate/gain tuning path from the audio engine. CrossfadeAudio still carried an applyTuning() method and its _playbackRate / _gainMultiplier fields — the pre-0.2.0 way of tuning a sound by nudging element playback rate and gain, superseded by real WAV regeneration and never called since. Both fields sat at their unity defaults (1), so every × _gainMultiplier was × 1 and every playbackRate = 1 was a no-op set of the element's own default. Removed the method, the fields, the now-unused applyTuning? member of the MixerSource interface, and the identity operations. Output is byte-for-byte identical — verified by the generator determinism tests and the live "audio flows through the shared master bus" check — this only tidies a superseded API out of the engine. Tagged in src/audio/sources.ts.

0.0.31

• Dev mode: the crescent moon holds still instead of drifting. The moon disc always ran a gentle breathing drift (moonDrift, ±6px), and dev mode layered its wane-to-crescent on top — so the dark dev-mode moon both darkened and floated. Dev mode now stops the drift (animation: none on .moon-dev), so it simply goes dark and stays put, as intended. Normal mode keeps its drifting moon; only the crescent-wane darkening is retained in dev. Tagged [v0.0.31] in index.css.

0.0.30

• Perf: the media-session player stops rebuilding on every volume tick. The single effect that sets the lock-screen / notification metadata and registers its play/pause/stop handlers depended on activeSounds (a fresh array each soundState change) and playAllActive (a fresh identity each change) — so dragging any volume slider rebuilt the MediaMetadata and re-registered all three action handlers on every frame of the drag. Split into three narrow effects: title/artwork update only when they actually change (the title is a string, so an unchanged active set is a no-op dependency), playbackState updates only on play/pause, and the transport handlers register once and read the latest functions through a ref. Same lock-screen behaviour, none of the per-drag churn. Tagged [v0.0.30 perf] in App.tsx.

0.0.29

• A11y: saving a mix is now announced. Deleting a mix already spoke a confirmation through the live region ("deleted mix …"), but saving one was silent — so a screen-reader user got no acknowledgement that the save succeeded, even though the visible UI (a new card, the field closing) told everyone else. handleSaveMix now announces "saved mix …", matching the delete path. Tagged [v0.0.29 a11y] in App.tsx.

0.0.28

• Mobile polish: the save-mix name field behaves on a phone keyboard. The "name this mix…" input now sets enterkeyhint="done" so the on-screen keyboard's action key reads "done" (Enter already saves the mix), and turns off autocorrect and spellcheck so a short name — often a proper noun like "Fireside" — isn't quietly rewritten or underlined mid-type. Desktop is unaffected. Tagged [v0.0.28] in MixControls.tsx.

0.0.27

• A11y: the save-mix name field now has a real label. The "name this mix…" input relied on its placeholder alone for a name — but a placeholder isn't a reliable accessible name (it disappears the moment you type and isn't consistently announced), so a screen-reader user focused on the field heard an unlabeled text box. Added an explicit aria-label ("Name this mix"). It's the app's only free-text input, and every slider/toggle was already labeled, so this closes the last unlabeled control. Tagged [v0.0.27 a11y] in MixControls.tsx.

0.0.26

• Fix: a stale scene deep link no longer strands you on a blank app. A ?scene=… launch (the app-icon shortcuts) that pointed at an unknown id — a scene since retired or renamed, or a typo — matched nothing, then returned early and skipped the "resume last night's mix" path, leaving an empty app. Now an unrecognised scene id falls through to the resume path (and the URL is cleaned either way), so a bad link still lands you on last night's mix instead of nothing. A valid scene still plays and returns as before. Tagged [v0.0.26 fix] in App.tsx.

0.0.25

• Fix: the greeting no longer goes stale. "good morning / afternoon / evening / up late" was recomputed only when App happened to re-render — so a phone left playing overnight (no timer running, nothing forcing a render) could still read "good evening" well past midnight, and coming back to a long-backgrounded tab showed yesterday's line. It's now held in state and refreshed on a one-minute cadence and whenever the tab returns to the foreground. Because setting the same string is a no-op in React, this only actually re-renders the handful of times a day the greeting truly changes — no idle churn. Tagged [v0.0.25 fix] in App.tsx.

0.0.24

• A11y: the sound card's editor toggle now exposes its open/closed state. The little "tune" button on an editable card opens and closes the inline sound editor, but carried only an aria-label — so a screen reader gave no hint of whether the panel was open, even though the editor's own "fine-tune sliders" disclosure already announces its state. Added aria-expanded to the card toggle so the two disclosures are consistent and the editor's state is legible without sight. No visual change. Tagged [v0.0.24 a11y] in SoundCard.tsx.

0.0.23

• A11y: the sleep-timer preset chips now read as durations to screen readers. The preset buttons showed terse text ("15m", "1h", "8h") with only aria-pressed — so a screen reader announced a bare "15m" (often spelled out), while the sibling extend/off buttons in the same row already carried full spoken labels ("Add 30 minutes to the sleep timer", "Turn the sleep timer off"). Each preset now has an aria-label ("Sleep timer, 15 minutes", …), so the whole timer row is consistently legible without sight. Visible chips and behaviour are unchanged. Tagged [v0.0.23 a11y] in MixControls.tsx.

0.0.22

• Fix: stopping the mix clears a running sleep timer. A stopped mix left its countdown frozen rather than cleared, and because the timer runs while playing, starting a fresh mix later silently resumed that leftover countdown — so a new session could stop itself early on a timer the user thought was long gone. The timer is now cleared whenever the mix empties (stop, drift-mode stop, or removing the last layer). It's a silent clear (stopping is already announced) and a no-op on timer expiry, and a scene/preset swap never triggers it — restoreMixerState replaces the layers in one batch, so the active set goes straight from old to new without passing through empty. clear() also gained a silent flag and now only speaks when a timer was actually set. Tagged [v0.0.22 fix] in App.tsx / useSleepTimer.ts.

0.0.21

• Perf: the scenes shelf and saved-mixes row no longer re-render on unrelated updates. Both rows were built inline in App, so every App re-render — once a second while a sleep timer counts down, and on each master/volume change — re-ran all ten scene cards and every saved-mix card, even though a card's own inputs (its scene/preset, whether it's the current mix, and play state) were unchanged. Extracted SceneCard and MixCard as memo'd components with ref-backed stable onPlay/onDelete handlers, so a timer tick repaints none of them — the same treatment SoundCard got in 0.0.11. No visual change. Tagged [v0.0.21 perf] in App.tsx.

0.0.20

• Fix: hand-tuning of WAV sounds now survives save and resume too. 0.0.16 taught the save paths to bake worklet layers' hand tuning; the WAV sounds (ocean, wind, chimes, and the rest that shape via variant chips) were still left out, so a mix saved — or fallen asleep to — with, say, a chosen ocean variant came back at the default render on reload. A new persist-time bakeForSave folds every enabled layer's effective editor tuning into the saved state: worklet layers unconditionally (params apply live, no cost), and WAV layers only when actually tuned away from their defaults, so an untouched WAV keeps using its memoized default render and never regenerates needlessly on load. Kept deliberately separate from the play-path enrichPresetState, which must not fold stale session editor values into a preset being loaded. Verified: a tuned WAV layer persists its tuning while an untouched one stays bare. No generator or scene changed. Tagged [v0.0.20 fix] in App.tsx.

0.0.19

• Perf: build the lock-screen artwork off the first-render path. The media-session "now playing" image (the crescent-moon night scene) was drawn to a 512×512 canvas and PNG-encoded inside a useMemo during the first render — a synchronous ~12 ms on this machine, more on a low-end phone — even though it isn't needed until the first play. It's now generated after first paint, on requestIdleCallback (falling back to a short timeout), so it no longer blocks the initial render; the media-session effect folds it in the instant it's ready, and the on-brand icon PNGs stand in until then. No visual change to the app; the lock-screen art is identical once built. Tagged [v0.0.19 perf] in App.tsx.

0.0.18

• Robustness: saving a mix can no longer throw. handleSaveMix minted the preset id with a bare crypto.randomUUID(), which is undefined in an insecure context (plain http on a non-localhost host) and in older browsers — so the save handler would throw and the mix silently fail to save. The storage layer already had a guarded safeId() (used by the migrations to backfill ids); it's now exported as newMixId and used here too, so the save path matches the migrations' never-throw contract. Tagged [v0.0.18 fix].

0.0.17

• Fix: the sleep timer now counts real time, not interval ticks. The countdown decremented by one each time its 1-second setInterval fired — but a sleep mixer runs with the screen off, exactly where browsers throttle (and locked phones suspend) background timers. A throttled tick can fire as little as once a minute, so the old timer ran many times too slow: a 30-minute timer could stretch into hours, and on returning to the app it simply resumed from where it had frozen, ignoring the real time that had passed. The timer now tracks a wall-clock deadline while playing and recomputes the remaining seconds from it on every tick and on every foreground return, so elapsed real time is always honoured and a backgrounded timer catches up the instant JS runs again. Pause still freezes it, resume re-derives the deadline, and extend pushes it out — all preserved. Verified with a controllable clock: fast- forwarding past the deadline stops the mix even when almost no ticks fire. Tagged [v0.0.17 fix] in useSleepTimer.ts.

0.0.16

• Fix: hand-tuning of worklet sounds now survives save and resume. Shaping a live worklet sound (rain, fire, birdsong, thunder, windy forest) in its editor stores the values in editorValuesBySound, which resets on reload — but the two persistence paths saved raw soundState, which never carries that tuning. So a mix you saved after shaping the rain, or the mix you fell asleep to, came back with its worklet layers at defaults. Both paths now persist the enriched state (enrichPresetState, the same baking already used when a preset plays), so what you saved is what you get back; editorValuesBySound is added to the resume effect's deps so a retune re-triggers the debounced save. Builds on 0.0.15 (which taught the writer to keep tuning) — that fixed scene-defined tuning, this fixes hand tuning. No generator or scene changed. Tagged [v0.0.16 fix] in App.tsx. (WAV-only sounds' hand tuning is still not persisted, matching preset semantics — a known, separate follow-up.)

0.0.15

• Fix: "resume your night" keeps each layer's character. The last-session writer (saveLastSession) persisted only { enabled, volume } per layer and dropped each layer's tuning. So reopening the app restored the layers of the mix you fell asleep to, but reverted every tuned layer to its defaults — e.g. a resumed Fan & Rain came back with a full default rain instead of the scene's quieter "at a window" bed under the fan. The writer now keeps a layer's tuning when present; the load path already sanitised and restored it (migrateSession → sanitizeTuning), so only this write side needed the fix. Added a migration test pinning that finite tuning survives (and junk / NaN is dropped). No generator or scene definition changed — this only makes the resumed mix faithfully reproduce what was already playing. Tagged [v0.0.15 fix] in savedMixes.ts.

0.0.14

• A11y: the snackbar no longer double-speaks. The forgiveness toast (stop mix / delete mix, with undo) carried role="status", making it a live region — but every action it shows is already announced by the app's dedicated live region ("deleted mix …") or its playback-status region ("stopped"). So a screen reader read each destructive action twice, in two slightly different wordings. The toast is now a plain container: not a live region (no double-announce) and deliberately not aria-hidden either (that would hide its own focusable undo button from assistive tech), so undo stays reachable. Its comment claimed this behaviour all along — the markup now matches. Tagged [v0.0.14 a11y] in Toast.tsx.

0.0.13

• Fix: the storage notice no longer covers the sleep timer. The one-time "we keep your mixes on this device" banner is a fixed element at the top of the z-stack (--z-toast), above both the now-playing sheet (--z-sheet) and drift mode (--z-modal). So on a first-ever play, opening the sheet put the banner squarely on top of the sleep-timer chips and fade note — a first-time user couldn't see or reach the timer. The notice is now held back while the sheet or drift mode is open (one prompt surface at a time, matching the app's existing prompt discipline) and reappears, still unacknowledged, the moment they close. Tagged [v0.0.13 fix] in App.tsx.

0.0.12

• A11y: the modal dialogs now trap keyboard focus. The now-playing sheet and drift mode are both aria-modal="true" and already moved focus in on open and restored it on close — but neither stopped Tab / Shift+Tab from walking out into the shell behind them, so a keyboard or screen-reader user could get lost outside an open modal while the rest of the page was meant to be inert. A small shared useFocusTrap hook (src/hooks/useFocusTrap.ts) now wraps focus at the dialog's edges and pulls it back if it lands outside. Purely additive — the existing focus-in/restore behaviour is untouched, and the trap is inert while the modal is closed. Tagged [v0.0.12 a11y].

0.0.11

• Perf: the sound grid no longer re-renders on unrelated updates. Every library card (SoundCard) received freshly-allocated inline callbacks each render, and handleSoundToggle changes identity whenever soundState does — so dragging one card's volume slider, or a once-per-second sleep-timer tick, re-rendered all ~19 cards. The card is now wrapped in memo, its callbacks are id-parameterized, and the parent passes ref-backed handlers with a constant identity. Result: only the card whose own props actually changed repaints; a timer tick repaints none of them. No visual or audio change — pure render-cost reduction, most noticeable while dragging a slider on a low-end phone. Tagged [v0.0.11 perf] in SoundCard.tsx and App.tsx.

0.0.10

• Revert the Fan stereo widening (0.0.9) — it sounded worse. The airflow-widening measured cleanly (correlation, mono-sum, band ratios all looked fine) but by ear the decorrelation thinned the whoosh and unsettled the steady hush that makes a fan calming. Reverted to the exact pre-0.0.9 mono render (verified byte-identical), and the generators test that pinned the fan as centred is restored, with a note recording why. The lesson: a clean measurement is not a listening test — the fan is a compact source and stays mono.

0.0.9

• Fan: stereo width on the airflow, tones still centred. A stereo-image audit (L/R correlation across the library) found the Fan the only broadband bed rendering fully mono — every other bed (ocean, wind, forest, train, airplane) is already widened. It had been kept mono on purpose, because decorrelating its near-pure blade/motor sinusoids combs them; but that also flattened the airflow "whoosh," which is the fan's largest and most genuinely diffuse component. Now the noise bed is split from the tones: the airflow above 900 Hz is decorrelated (L/R correlation 1.00 → 0.72, matching the airplane bed) while the blade pulse, motor hum, casing resonances and all bass stay dead-centre — width without combing the tones. The mono sum is nearly unchanged (all bass fused), so it still sounds like a fan on a phone speaker; on headphones it fills the room. Verified by A/B against the old render and through the live audio graph.

0.0.8

• WAV rendering moved off the main thread (Web Worker). Every sound's first play and every variant/slider retune renders 32 s of audio — a synchronous 100 ms–3 s DSP block (ocean ~670 ms, fire ~560 ms on this machine) that ran on the main thread, freezing the UI mid-tap so even the card's loading spinner couldn't animate. Rendering now runs in a dedicated worker (src/audio/genWorker.ts); the blob URL is minted in the same origin's store, so only the URL string crosses back and the <audio> elements play it directly. Falls back to the old inline render if a worker can't start (CSP, test runners) — same deterministic output. Verified bit-for-bit parity with the main-thread path and 19/19 through the live audio graph.
• Hardened the sound-sweep verifier against sparse sounds. The permanent runtime sweep read a single ~6 ms peak-meter window per 250 ms poll, which almost never coincides with a sparse transient — a clock tick is ~30 ms once a second — so it could report the (perfectly audible) Ticking Clock as a FAIL. It now accumulates the peak per animation frame across each listen window, catching every tick; confirmed the clock ticks nine to eleven times in eight seconds at ~0.3–0.4 peak.

0.0.7

• Scene balance audit: every advertised layer is now actually audible. Scene volumes were tuned long before months of generator re-leveling, and a per-layer measurement (WAV rms at the scene's tuning × volume × masking gain, in dB against the scene's lead) showed several scenes no longer delivering their own mood line: Fan & Rain's fan — the title layer — sat at −20.5 dB (the scene played as a Rainfall duplicate; now −5.4 dB, a proper co-headliner), Fireside's "quiet dark" insects were at −22.4 dB (now −12.3, a real garnish), Deep Rest's heartbeat was masked at −10.5 dB inside the same low band as the brown noise (now −6.5) with its insects at −16 (now −11.6), and Ocean Night's insects were at −16.7 (now −10.7). Targets: co-headline −3..−7 dB, named garnish −9..−13 dB, nothing advertised below −15. Mix numbers only — no generator changed.

0.0.6

• Ticking Clock: audible at last. A whole-library level audit found the clock ~20 dB under everything else (rms 0.011 vs the library's 0.08–0.16) — as an accent layer it gets no masking compensation, so in any mix it simply vanished, and solo it was startlingly quiet next to any other sound at the same slider. Fixed without changing the approved bare-tick character: the same sharp attack now carries a quiet ~11 ms gear-train tail, and the sparse transients normalize to 0.8 peak (they read far quieter than continuous beds at equal peak). Result: +7 dB energy across all four variants, true silence between beats preserved (floor at the dither level), and the clock is still the quietest sound in the library by sustained energy — present, not prominent. Verified per-variant by probe and through the live audio graph (19/19 sweep).

0.0.5

• Descriptive variant marks for the last six bar-only sounds. Birdsong, Stream, Thunder, Night Insects, Underwater, and Shower were still using the generic ascending-bars fallback on their character chips; each now has crafted 16px marks in the house grammar (lib/variantIcons.tsx, inline SVG — no icon-font subset step involved): far song / a perched bird / the dawn sun; a trickle / stones in a brook / rushing flow; a flicker beyond the horizon / rumble under a cloud / a gathering second cell / the bolt landing; a lone chirp in the grass / the moon over a meadow / deep-night stars; hanging bubbles / a current / the deep floor; and three shower pressures. Every sound with variants now has descriptive icons — the bars remain only as the deliberate fallback for future unmarked variants.

0.0.4

• Fix: leftover "drift" branding in the install prompt. The install row's aria-label still read "Install drift" from before the rebrand; now "Install starlight." Found while investigating a report of the install prompt not appearing (root cause was the intentional hasPlayed && storageAck gate plus per-browser support, not a bug — see InstallPrompt.tsx).

0.0.3

• Privacy page: dropped the GitHub-repo link. The closing "questions? open an issue" note pointed at the repo; removed along with its now-unused .note styling (the border/spacing moved onto the "back to starlight" link so the page's closing rhythm is unchanged).

0.0.2

• Custom domain: the app now lives at https://starlight.rest/. GitHub Pages 301-redirects the old hamzaanjum96.github.io/SleepSounds/ URLs to the apex of the new domain — which broke the app there, because the build was rooted at /SleepSounds/ and every asset 404'd after the redirect stripped the repo path. Fixes:
  • Vite base is now / (the custom domain serves from the root); the Playwright/e2e, sound-sweep, and OG-card scripts follow.
  • Every absolute URL — canonical, og:url, og:image, Twitter image, JSON-LD, robots.txt's sitemap pointer, sitemap.xml, llms.txt, the privacy page's canonical — now points at https://starlight.rest/.
  • Old links keep working via GitHub's redirect; HTTPS on the new domain activates as soon as GitHub finishes provisioning the certificate (settings-side, nothing in-repo).

0.0.1

• Rebrand: the app is now "starlight" (formerly drift away), and the version counter restarts at 0.0.1 for the new brand — every entry below this one is the drift-away era. The night theme already fit the name, so this is a text/metadata pass:
  • In-app: wordmark, document title, media-session title/artist, storage notice, crash line.
  • Metadata: page meta + OG/Twitter tags, JSON-LD, PWA manifest name/short name, privacy page, robots.txt, llms.txt.
  • Social card: regenerated with the starlight wordmark and shipped under a fresh URL (og-starlight.jpg) so WhatsApp's cached preview of the old brand can't linger; the old og-card.jpg is gone.
  • Service worker: cache prefix renamed to starlight-; activation still deletes old drift-away-* caches so upgrading devices clean up.
  • Docs: README/ARCHITECTURE/DESIGN/ADDING-A-SOUND intros; the package is named starlight. Historical changelog entries keep the old name. Internal identifiers (localStorage keys, __drift* globals, .drift-slider classes, drift mode the feature) are deliberately unchanged — renaming storage keys would discard users' saved mixes.

9.5.4

• Privacy page metadata. The privacy page is listed in the sitemap but had no meta description or canonical URL; search results would improvise a snippet from whatever text they found. It now carries both, with a description that mirrors what the page actually promises (no accounts, no analytics, nothing leaves the device).

9.5.3

• Documentation refresh + one honest number. The docs pass, applied to the docs themselves:
  • The public sound count now tells the truth everywhere: 19 sounds are in the lineup (Stream and Shower are pulled — HIDDEN_SOUND_IDS), but the OG/Twitter descriptions, JSON-LD, llms.txt, the README, and the in-app library counter all said 21. All now say 19, and the in-app counter is computed from the visible library (so dev mode counts what it actually shows).
  • README: maintainer notes now cover the two new invariants (keep the discoverability surfaces in sync; UI changes must pass the axe gate), a Discoverability section documents the SEO/AI-reader surfaces in one place, developer tips document dev mode and the test commands, and the features list reflects the character-preset editor system.
  • DESIGN.md's accessibility section records the landmark structure and the axe-core gate added in 9.5.2.

9.5.2

• Accessibility: an automated axe-core gate, and landmark fixes.
  • New e2e spec (tests/e2e/a11y.spec.ts, via @axe-core/playwright): no serious or critical violations allowed on the main page, the now-playing sheet, an open sound editor, or the desktop split layout — so label/ARIA/contrast regressions can't creep in silently.
  • Landmarks: the app's content now sits in a real <main>; the footer is a <footer> element; the hidden site-verification text is aria-hidden. The page now scans clean at every axe impact level, not just the gated ones.

9.5.1

• llms.txt for AI readers. A plain-text, agent-oriented summary at the site root (per the emerging llms.txt convention): what the app is, the full feature list, the key facts (free, open source, offline, no tracking), and links to the in-repo docs. It notes explicitly that the app is a client-side React shell — fetching the HTML yields nothing — so this file and the JSON-LD are the reliable machine-readable descriptions.
• Fix: removed an older, second JSON-LD WebApplication block that predated 9.5.0 — two entities describing the same app confuse structured-data parsers; the richer 9.5.0 block is the one that stays.

9.5.0

• SEO fundamentals. Start of a discoverability/accessibility pass:
  • JSON-LD structured data (schema.org/WebApplication) in the page head — name, rich description, category, price (free), feature list, card image and screenshot — the machine-readable summary search engines and AI readers prefer over scraping.
  • robots.txt (everyone welcome, AI crawlers included — the app is fully client-side, there's nothing to protect) pointing at a new sitemap.xml (the app + the privacy page).
  • Verified: JSON-LD parses in the built output; both files ship in dist/.

9.4.5

• Train: the static layer is removed, not just turned down. 9.1.6 halved the hiss bands; feedback says halving wasn't the point. The wheel-top (2.3–5.2 kHz) and aero (1.5–6.4 kHz) noise sources are now deleted — at sleep volumes they only ever read as static — and the surviving mid floor (rolling band, rail mid) is darker again and rides the carriage sway and roughness flutter, so nothing in the sound holds one frozen level. A sleeping carriage is rumble, sway, and joint clacks; any brightness left belongs to the clatter events. Measured: energy above 1.5 kHz falls to 1.1–1.9% across all four scenes (Express was 8.3%).

9.4.4

• Underwater: the laser guns are gone — full revert of the bubble experiment. The 9.3.8 softening wasn't enough; the physics voicing was the problem itself. Its exponentially decaying bubbles ring as pure tones for up to ~0.9 s (a 150 Hz "glug" has a long ring time by design), and an exposed pure-tone burst against this dark muffled bed reads as weapon fire no matter how gently it starts. The bubbles are back to the pre-9.3.0 voice that was right all along: short soft blips (8–30 ms, sine envelope, mild rise), glug trains deleted. The bubble-physics notes stay in docs/research/ — the stream's babble keeps that model, where it works.

9.4.3

• Deploy hardening, round two. 9.4.2's Pages deploy was rejected twice — the 90 s retry wasn't enough, so the backend's bad mood can outlast a short backoff. The deploy job now makes a third attempt after a 5-minute pause, and the workflow gained a workflow_dispatch trigger so a rejected deploy can be re-run from the Actions tab without cutting a release (it always ships the latest main). This release's own push also redelivers 9.4.2's fixes (the click-dead side panel and the moon anchor) to the live site.

9.4.2

• Two desktop bugs: the click-dead side panel, and the wandering moon.
  • Side panel sometimes unclickable (including play): the bg-layer gradient wash — a fixed, viewport-covering, positioned element — was the one scenery layer without pointer-events: none, and the side panel's content was unpositioned, so the invisible background hit-tested above the entire panel and swallowed every click (reproduced: elementFromPoint over the panel's play button returned bg-layer). Fixed both ways: bg-layer is now pointer-transparent like the other scenery, and the panel joined the app shell's stacking layer so no scenery can ever sit over it again.
  • Moon sliding down as the page grows: the compositor scroll-parallax mapped progress over the total scroll height, so the same scroll offset put the moon lower on longer pages (the JS fallback was pixel-based and didn't drift — the two paths disagreed). The animation now runs over a fixed range (animation-range: 0 100vh, the same 0.7px-per-px as the fallback) and the resting anchor is top: 7vh — viewport units, so nothing about page length can move it. Verified: at the same scroll offset the moon now sits at the identical pixel on a short and a lengthened page.

9.4.1

• Windy Forest, round two: less harsh, and no longer static-like. The static character was the constant grain spray — leaves fizzing at a steady base rate regardless of wind. Rustle is now gust-carried: the base rate drops ~60% so lulls genuinely rest, and the density lives in the gust term, arriving as waves of rustle with real pauses. The whole voice also sits lower: leaf grains at 700–2600 Hz (another notch down), the wind's body band leads while the mids and whistle recede, the master lowpass comes down to ~3.4 kHz, and the WAV fallback's bed and twigs are darkened to match.

9.4.0

• A proper social-share card. Sharing the link (WhatsApp, iMessage, Slack, X…) now previews a generated 1200×630 night scene — the app's textured starfield, the moon, and drift away in the Cormorant wordmark with the line "sleep sounds, generated in your browser" — instead of the square app icon in a small card. Built from the app's own visual ingredients by scripts/og-card/render-og-card.mjs (generated, not borrowed), shipped as a 46 KB progressive JPEG (well inside WhatsApp's preview size limit), with refreshed og: / twitter: meta: summary_large_image, absolute image URLs (WhatsApp requires them), proper dimensions/alt text, and a description that says what the app actually is — 21 sounds, generated in the browser, free, offline, no tracking. Note: WhatsApp caches link previews aggressively; a fresh image URL (og-card.jpg) sidesteps the stale cache.

9.3.11

• Heartbeat: no more keyboard-click on the beat. The blood-rush layer was fine; the beat itself ticked. Each thump mixed in raw broadband white noise ("tissue," at 0.22), which put a wideband transient spike up to ~2 kHz on top of every low beat — exactly a keyboard-key click on a spectrogram. Real heart sounds are low-frequency, so the murmur is now heavily lowpassed (~180 Hz) and much quieter (felt as body, not heard as a tick), and the thump onset is softened (15 ms → 24 ms rise). Measured: the beat's spectral centroid drops from 382 Hz to 128 Hz and its energy above 800 Hz from 9% to 2.5% — the bright click collapses into the thump.

9.3.10

• Windy Forest is softer. The harshness was a bright, hissy top: leaf grains reached to 6 kHz, the whistle band was resonant (Q 4), and nothing tamed the top end. Leaf rustle now sits ~900–3200 Hz (the old top octave was the sizzle), each grain a touch quieter; the whistle band is eased (Q 4 → 3, gain 0.5 → 0.38); and a gentle master lowpass (~4.8 kHz) rolls off the canopy's top so it reads as soft, not brittle. The WAV fallback's twigs were darkened to match (top cut 5.4 → 3.8 kHz). (This pass also added the missing lowpass method to the worklet's filter class — the master lowpass needs it; caught before shipping because the runtime sweep flagged the forest going silent.)

9.3.9

• Thunder: the static bed is gone, and the sliders are back. The live worklet already falls to true silence between strikes (measured: 53% of the loop near-zero), so the "static white noise bed" was the WAV fallback (served when a device can't load the worklet): it ran a continuous brown-noise rumble the whole time. That bed is now gated by a per-strike bloom envelope — the roll swells with each strike and rolls away over several seconds (longer for far storms), fading to silence before the next, matching the worklet's "rolls, then quiet" shape (fallback now measures 54% near-silent, like the worklet). Removed the fallback's leftover pink-noise hiss band in the same pass. And the three sliders — activity / rumble / distance — are no longer hidden (variantsOnly is off), so you can shape a scene, not just pick one.

9.3.8

• Underwater: no more laser guns. Feedback nailed it — an exposed pure sine chirping 35% from a 2 ms edge reads as a weapon, not a bubble. The bubble voice keeps its physics but is now heard the way water actually delivers it: a ~12 ms eased onset (water blurs attacks), only a hint of up-chirp (8%), lower frequencies (300–900 Hz), roughly half the level, and a sparser population. The glug trains stay, at less than half their old weight.

9.3.7

• Thunder's WAV fallback opens gently, like the real one. The live worklet holds its first strike back 9–21 s and softens it — the sleep-safe opening from the 7.19 rework — but the fallback loop could land a full boom 1.5 s after you pressed play. It now opens with the same lull (8–15 s, loop-safe: on repeat it reads as a pause between strikes) and its first boom lands at 60% weight.

9.3.6

• Windy Forest's WAV fallback comes off its rails. The live worklet is already a proper gust-field model; the fallback loop (used when the worklet can't load) still swelled on a fixed-frequency |sin| (an audible ~15–30 s cycle) with twigs cracking on a metronome regardless of the wind. The canopy now breathes on a loop-closed random gust walk and the twigs ride it — mostly cracking when the canopy is actually moving, harder in strong gusts, mostly silent in lulls.

9.3.5

• Brown Noise: a slow, loop-closed swell (±10%, 5–11 s holds) so the deep floor rolls almost imperceptibly instead of holding one frozen level — brown is all low end, so it breathes in level where white/pink breathe in tilt.

9.3.4

• Pink Noise: the same gentle spectral breathing as 9.3.3's white — a loop-closed walk toward a darker copy of itself, so the fall stays uneventful without holding one frozen spectrum for hours.

9.3.3

• White Noise: the spectrum breathes. Auditory-texture work says realism lives in slowly varying subband statistics, and perfectly stationary noise reads as synthetic and fatiguing over hours. The body now crossfades very gently toward a darker copy of itself on a loop-closed random walk (spectral tilt breathing, never a level pump) on top of the existing drift/shimmer movement. Uneventful, but no longer frozen.

9.3.2

• Shower: water strikes a surface now. It was three noise bands — a vent, not a shower. Following the "sound atoms" split from the rain research (background bed + discrete impacts), a dense patter of individual drop impacts now sits under the spray, its rate riding the pressure slider, and the occasional heavier hit rings the tub with a fast-damped low resonance. Spray, steam and the room resonance are unchanged. Hidden from the lineup for now — audition it in dev mode.

9.3.1

• Stream actually babbles. The generator was two filtered noise bands — which is precisely the "indistinct murmur" failure the liquid-sound literature describes; a brook's voice is its individual bubble events. A babble layer now rides the flow: a population of damped-sinusoid bubbles with the rising Minnaert chirp (mid-band voices, with the sparkle slider governing the bright small-bubble end), each panned to its own spot, their rate surging with the flow. A quiet collective low band (the coupled-bubble hum of dense water) rides the depth slider. Hidden from the lineup for now — audition it in dev mode; it's a candidate to return.

9.3.0

• Underwater: bubbles that obey bubble physics. Start of a research-grounded pass over the less-worked sounds (notes + sources in docs/research/water-family-synthesis.md). Per the Minnaert / van den Doel liquid-sound model, a real bubble is a damped sinusoid with a near-instant onset, an exponential decay that lengthens with bubble size, and a pitch that rises through the decay — the old bubbles were symmetric sine-envelope "boops" with none of that, which is exactly the electronic blip the literature warns about. The population is Minnaert-spread too: mostly small distant bubbles heard darkly through the water, with occasional low glug trains (a large bubble breaking into 2–4 smaller, higher ones as it rises). The deep pressure bed and current swell are unchanged.

9.2.6

• Pages deploys now retry a backend rejection. 9.2.4's queue-don't-cancel fix wasn't enough: its own deploy run failed the same way ("Deployment failed, try again later" straight from the Pages API) even without any cancellation in play — the backend just intermittently rejects deployments when several land minutes apart. The deploy job now treats the first attempt as fallible, pauses 90 s, and deploys once more for real before the run is allowed to fail. Between queueing (9.2.4) and this retry, a burst of releases can no longer strand the site on an older build.

9.2.5

• The library is ordered by build maturity. The grid now leads with the most convincing synths — Rain and Fire (the flagship worklets), then the long-settled Fan / Birdsong / Night Insects, then Thunder, Windy Forest, Ocean, Wind, Underwater, the noise family, Train, Airplane, and the newest builds (Heartbeat, Cat Purr, Wind Chimes, Ticking Clock) last. The hidden sounds close the list, so dev mode appends them at the end.

9.2.4

• Fixed the failing GitHub Pages deploys. The 9.2.0 and 9.2.1 deploy runs died with "Deployment failed, try again later" while CI stayed green. Cause: the deploy workflow used cancel-in-progress: true, and cancelling a Pages run mid-deployment leaves its server-side deployment dangling — the next run then collides with it. With releases landing minutes apart, that happened back to back. Deploys now queue instead of cancelling (cancel-in-progress: false, per GitHub's own Pages guidance for production deployments). The site itself was never stale for long — each successful deploy publishes the latest main.

9.2.3

• Category filters reordered. The elemental families people reach for first lead: Water · Fire · Air · Wildlife, then the indoor/textural ones (Cozy · Earth · Noise · Urban).

9.2.2

• Cat Purr moved to Wildlife. It's an animal, not furniture — it now sits with Birdsong and Night Insects (and wears the moss accent), leaving Cozy to the hearth-side textures: Heartbeat, Wind Chimes, Ticking Clock.

9.2.1

• Dev mode: spam-tap the moon. Five quick taps on the moon (within 3 s) toggle dev mode; five more — or a refresh — turn it off (session-only by design, nothing persisted). While it's on, the moon wanes to a crescent (a night-sky disc slides across, the glow pulls in), the greeting reads "… · in dev mode", and the library reveals the hidden / pulled sounds (Stream, Shower today — anything in HIDDEN_SOUND_IDS plus experimental builds). Plumbing: the moon is a real (a11y-hidden, tab-unreachable) button now, and it needed its own z-index step above the app shell — real taps hit the shell, not the scenery, which is also why the old "moon can't swallow taps" rule still holds: only the 52 px disc itself catches anything.

9.2.0

• Ticking Clock: the bare tick is the sound now — and the sliders vary the mechanism, not the mud. Feedback was consistent: it only sounded right with every slider at zero. So that render — the naked escapement click, true silence between beats — is the generator's base voice, and the case knock / distance muffle / room tone controls are gone entirely. The new sliders edit real qualities of the mechanism: pace (a slow 2-second pendulum through the 1-second default to a quick ½-second pocket-watch tick, always snapped to an even beat count so the tick/tock alternation survives the loop seam), tick–tock (how differently the two escapement faces are voiced, from one even tick to strongly two-toned), and brightness (the click's tone). New chips with new marks: Slow Pendulum · Bare Tick (default) · Even Beat · Pocket Watch. Measured: default = 32 beats and 93% true silence; Slow Pendulum 18 beats; Pocket Watch 60.

9.1.7

• Every variant that depicts something now has a drawn mark. The crafted 16px chip marks (previously only rain's roof/window/tin and fire's flames) now cover ten sounds — 39 new marks: ocean's ripples / horizon surf / cresting roller / storm spray; wind's starlit streamline / hillside / roof eaves / stacked gale; the four fans as their actual appliances; the woods (trembling leaves, tree crown, twin pines, a pine bent into the wind); train's receding rails / lit sleeper carriage / jointed rail / express nose; airplane's crescent-and-plane / level cruise / swept wing / bumpy flight line; heartbeat's drifting heart / heart / filled heart / heart held in a circle (Womb); the clock as pendulum / mantel arch / tall case / tick-through-a-doorway; the chime sets hanging still, long, leaning, and thrown wide; and three cat faces for the purr (eyes closed, settled, rumble radiating). All inherit the chip's category accent via currentColor; the noise family keeps position bars deliberately — a spectral tier is that family's honest character axis.

9.1.6

• Train no longer sounds like overpowering static. The continuous floor's broadband bands (rolling noise, rail mid, wheel top, aero hiss) sat on top of the rumble instead of under it. The rolling band is darker (lowpassed ~30% lower with a second pole so it can never read as open hiss), the rail mid / wheel top / aero weights are roughly halved, and the underfloor body leads. Measured on the default Sleeper Car: energy above 1.5 kHz drops from 11% to 3.7%, the 0.3–1.5 kHz band from 16% to 10% — the carriage now reads as rumble and joint clacks, not static.

9.1.5

• Windy Forest no longer slams in — and every sound now eases in. Three layers to the harsh kick-off: the forest worklet woke at full wind (its wind-speed state initialised at the mean, its internal gate ramped in 80 ms) — it now wakes in a lull (wind state starts low and climbs over a few seconds, the gust detector starts settled, the gate ramps over ~0.5 s) so the wind arrives instead of appearing. And the mixer's fades are now asymmetric: fade-in lengthened from 700 ms to a 1.6 s quadratic ease (the first moments swell rather than step — this is what also soothed Ocean's kick-off), while fade-out stays at 700 ms so stopping remains responsive. Measured onset at the master bus: a smooth build over ~3 s for both forest and ocean, no step.

9.1.4

• Ocean calmed — the gentle shores are now actually gentle. Every variant played at roughly the same loudness because the WAV encode is peak-normalised (a calm render just got boosted back up); the output gain now follows the scene's intensity, so Lapping Shore plays ~3 dB under Storm Surf instead of matching it. The break rides foam squared (at low foam the crest folds almost soundlessly), its attack eased from 120 ms to 200 ms (a fold-over, not a slap), its top darkened, the surf wash scales with wave size, and the backwash sits lower. Measured: calm scenes now carry 3–4% of their energy above 800 Hz vs the storm's 12% (previously all variants were equally bright and equally loud). Lapping Shore and Distant Surf variant params also ease their foam further.

9.1.3

• Cat Purr breathes like a sleeping cat, not a panting one. The breath cycle ran ~2.1 s (≈28 breaths/min) — a listener entrains to a breath rate, and that fast reads as anxious. The default is now ~3.4 s per breath (measured: ≈17/min across the loop, still varying breath to breath); the breathing slider keeps its range, just mapped slower end to end.

9.1.2

• The runtime sound sweep is now a tool, not a one-off. scripts/sweep-sounds.mjs drives the real app headless, toggles every library sound, and asserts signal reaches the master bus — the end-to-end check that caught 9.1.1's buried Birdsong while every unit test stayed green. Listen windows are sized per sound (Thunder's deliberate 9–21 s quiet opening gets 30 s).
• docs/ADDING-A-SOUND.md — the full checklist for adding a generated sound: the synthesis rules the library lives by (seeded determinism, loop-clean structure, no fixed-frequency LFOs, resonator recurrences over per-sample sin·exp, stereo-with-intent), the seven wiring points and what breaks if each is missed, the verification recipe (numeric probes → spectrograms → runtime sweep), and the variant/label conventions. Referenced from the README tree and ARCHITECTURE.md.
• Variant chips: every mark is distinct now. Position bars were capped at three, so 4-variant sounds drew their last two chips with identical marks; the bar set now sizes itself to the variant count (a 4-chip set draws four ascending bars).

9.1.1

• Bug fix: live Birdsong was near-silent — for months. A full-library runtime sweep (play every sound, assert signal at the master bus) caught it: the worklet's output "click smoother" (prev += 0.03·(x−prev) per sample) is mathematically a one-pole lowpass at ~210 Hz, applied to the whole signal — and birdsong lives at 2–6 kHz, so the live sound was buried ~20 dB (peak 0.017 where its own WAV fallback renders 0.93; an earlier "birdsong is quiet" ×1.5 lift had compensated the symptom without finding this). The fix smooths the run/stop gate, not the signal, and level-matches the worklet to its fallback (long-run RMS 0.084 vs 0.085 at identical default params) — so failing over is seamless and the 0.34 default volume means the same loudness on either route. In-app: first call now registers at ~5 s where the old build showed zero signal for 25 s. Also confirmed no other worklet shares the pattern (fire's one-pole is its intentional roar-brightness filter), and Thunder's silent opening is by design (first strike held back 9–21 s).

9.1.0 — Cleanup, docs & cache-correctness milestone

A maintenance release closing out the 9.0.x sound pass.

• Bug-class fix: in-place public/ asset updates now reach installed clients. The SW build id previously hashed only the precached file names, so any stable-named asset edited in place — a re-subset icon font, and notably every audio worklet (rain, fire, thunder, birdsong, windy forest all shipped in-place retunes over their history) — could leave installed clients running old bytes. The build id now hashes every precached file's contents (scripts/inject-precache.mjs), and the SW install precaches with cache: 'reload' requests so a new build can't re-cache stale bytes out of the HTTP cache (public/sw.js). Verified: a single flipped byte in the icon font changes the build id.
• The icon-subset process is now documented in three places: a full docstring in scripts/subset-icons.py (what a ligature font is, when to re-run, the failure mode — a missing glyph renders as its raw name in text — and the cache-safety story), a new "Icons: the Material Symbols subset" section in this README, and a "Fonts & icons" section in ARCHITECTURE.md.
• Dead code removed. layeringTrim (superseded by layerShaping, which the mixer actually uses — its invariant tests now target layerShaping directly) and the never-consumed defaultWidth field on every LAYER_META entry (stereo width has been baked into generation since v4).
• Docs de-drifted. The maintenance note at the top of this README no longer tells you to bump the saved-mixes key for generation changes (loops re-render every session; bumping would discard user mixes — it's only for incompatible param-meaning changes), DESIGN.md no longer references the removed layeringTrim, and ARCHITECTURE.md documents the PWA cache-correctness rules above.

9.0.13

• Brown Noise: character chips. Smooth / Rolling / Deep was a depth dial; the set is now Velvet (a featureless, fully smoothed floor) · Rolling (default) · Storm Floor (heavy, grainy low end). Sliders were already honest (depth / rumble / smoothness) and stay.

9.0.12

• Pink Noise: honest labels, real characters. "focus" is labeled presence (it brings the midrange band forward — that's all it ever did). Chips: Warm Blanket (pulled over your head) · Soft Fall (default) · Mountain Air (thin and bright, mids stepping forward).

9.0.11

• White Noise: the shimmer came off its rails, and the labels tell the truth. The air band's movement was a fixed 0.065 Hz sine — an audible ~15-second cycle in a sound whose whole job is to be uneventful; it's now a loop-closed random walk. "depth" is labeled body (it sets the low floor), "texture" is shimmer (it sets how much the air band moves). Four characters: Warm Hush (dark masking veil) · Even Veil (default) · Open Air (living, airy) · Crisp (bright full-range static).

9.0.10

• Airplane chips are seats and moments. Cabin / Cruise / Turbulent was a dial; the set is now Night Flight (the hushed rear cabin, vents leading, almost no rough air) · Cruise (default) · Over the Wing (the boundary-layer roar and engines lead) · Light Chop (a stretch of gentle rough air). The "cabin" slider is labeled cabin hush — it always controlled the ventilation bed, not the engines.

9.0.9

• Train chips are journeys. Distant Rails / Steady Carriage / Fast Express read as a speed dial; the set is now Distant Line (a railway heard across the fields, hardly any joint noise) · Sleeper Car (the default overnight carriage) · Old Local (slow, heavy, riding jointed track — the clackety character) · Express (fast and smooth). The "clatter" slider is labeled rail clatter.

9.0.8

• Windy Forest chips are kinds of woods. Light Rustle / Breezy Canopy / Storm in the Trees was a wind-speed ramp; the set is now Aspen Shimmer (leaves trembling in barely any wind) · Breezy Canopy (default) · Deep Woods (close old-growth, branch creak and twig detail leading) · Before the Storm (the whole forest heaving). The presets drive the live AudioWorklet and the WAV fallback identically.

9.0.7

• Fan chips are appliances now. The synthesis already models fan acoustics properly (dual airflow bands, casing resonances, blade-amplitude drift from a research pass), so this release replaces the Low Hum / Steady / High Speed intensity tiers with machines you'd actually put in a room: Air Purifier (small, smooth, nearly toneless hiss) · Bedroom Fan (the default) · Box Fan (fuller drone, more hum) · Shop Fan (big, slow-bladed, room-moving air). Sliders were already concrete (speed / hum / airflow / fan size) and stay.

9.0.6

• Wind gusts no longer run on rails. The gusting was three fixed sine waves multiplied together, so the same swell pattern repeated every ~26 seconds — exactly the kind of cycle a half-asleep ear latches onto. Gusts now arrive on irregular, loop-closed random walks with fast micro-flutter inside them. The edge-tone whistles also behave physically now: they only sing when the wind is actually up (gated by the gust level) and their pitch follows the brightness slider (renamed from the vague "tone"). Variants are places, not levels: Night Breeze · Open Hillside · Around the Eaves (whistle-led, wind heard around a building) · Winter Gale. The Evening Porch scene's wind layer now matches the Night Breeze variant exactly, so its editor shows the name.

9.0.5

• Ocean waves now break. The old wave was one noise band swelling and fading — surf with no surf. Each wave now carries the full shoreline anatomy: the undertow swell builds, the crest breaks (a brighter burst that rises in ~120 ms of real time as the wave folds over, then dies across the wash), and a low granular backwash rakes back down the slope before the next wave. The break rides the crash & foam slider (renamed from "foam"); "depth" is now labeled undertow, which is what it always was. Variants become shore scenes, calm → wild: Lapping Shore (small quick laps), Distant Surf (big waves heard from far up the beach), Rolling Waves and Storm Surf.

9.0.4

• Heartbeat rebuilt with real cardiac timing — and a womb. The old beat was two fixed sine blips 200 ms apart. Now S1 ("lub") opens systole and a softer, higher S2 ("dub") lands about a third of the cycle later, leaving the long diastolic rest that makes a heart read as slow even at the same BPM; each sound is a pressure thump whose pitch falls as it decays (a valve closing into tissue, not a note), and beat timing/weight vary a little every cycle. A new blood flow slider gates a circulatory rush to the cycle — raised together with the muffle it becomes the classic womb sound. Sliders: pace · chest depth · muffle · blood flow. Variants are characters, not tiers: Falling Asleep · Resting · Against the Chest · Womb.

9.0.3

• Ticking Clock rebuilt around the bare click. The feedback was blunt: the only passable setting was every slider at zero — i.e. the naked click. The wooden case "thump" was a pure sine (a marimba note, not a clock case) and the distance filter muddied everything above it. The case is now a knock — a 4 ms noise burst ringing a fast-damped resonance, so it thuds like wood instead of singing — the distance range keeps zero genuinely crisp, the room-tone floor is darker and quieter, and defaults pull everything back toward the mechanism (wood 0.45, distance 0.45, room 0.25). The character the feedback liked is now a chip: Bare Tick (90% of its loop is true silence). The set: Bare Tick · Mantel · Grandfather · Distant Hall, and the "wood" slider is labeled wood knock.

9.0.2

• Wind Chimes: the static bed is gone. The faint pink-noise "breeze" that ran under the strikes read as hiss, so it's removed outright — between gusts the chimes now rest in true silence (~11% of the loop measures fully silent). The orphaned "air" slider became ring: it sweeps the tubes' sustain from damped to long-tolling. Variants replace the 3-step intensity ramp with four characters: Still Evening (a rare dark stir), Deep Tubes (long low tolls), On a Breeze (the porch default) and Dancing (bright and lively).

9.0.1

• Cat Purr no longer sounds mechanical. The first cut repeated one identical breath for the whole loop — same length, same pitch, same weight — which is exactly what a machine does and a cat doesn't. Now every breath is drawn individually (length ±8%, pitch ±0.8 Hz, weight ±8%, exhale share ±3%, all summing exactly to the loop), the pitch settles ~5% as each exhale runs out, pulse timing jitter is wider, and a slow unsynchronised sway sits under everything — the cat shifting its weight, not a machine holding a level. Verified: breath spacing now spreads 0.76–1.37 s across the loop (was constant), spectrum unchanged (~97% of energy below 200 Hz).

9.0.0

• Fixed the missing icons for Cat Purr / Wind Chimes / Ticking Clock on already-installed clients. The 8.3.0 icon-font re-subset kept the file name material-7.woff2, and the service worker's cache version hashes the asset names, not their bytes — so installed clients kept serving the old cached font and the three new glyphs (paw, bell, clock) rendered as placeholder text. The font now ships as material-8.woff2, which busts the SW precache and any HTTP cache in one move. Starting the 9.x line here; the sound- quality pass that follows ships one sound per patch release.

8.3.0

• Three new generated sounds: Cat Purr, Wind Chimes, Ticking Clock. The Cozy family grows from one sound to four, all procedural (nothing sampled):
  • Cat Purr — a ~25 Hz glottal pulse train that actually breathes: a louder, lower exhale phase, a turnaround pause, then a softer, higher inhale, with the breath period snapped to a whole number of cycles per loop. Grains ring chest/throat resonances with eased onsets (no rasp), a faint breath-noise band rides the envelope, and a double lowpass keeps ~97% of the energy below 200 Hz. Editors: breathing / rumble / softness; variants Dozing · Content · Deep Rumble.
  • Wind Chimes — five low pentatonic tubes (A3–F#4) struck in gust-driven clusters, each strike ringing the first four transverse modes of a free tube (1 : 2.76 : 5.40 : 8.93, higher modes decaying faster) via damped- resonator recurrences (~10× faster than sin·exp per sample; renders in line with rain/ocean). A strike often swings on into a neighbouring tube; lulls are bridged by a forced lone soft strike so the loop never goes dead for ten straight seconds, and a faint decorrelated breeze bed fills the air between. Stereo, each tube at its own place in the field. Editors: breeze strength / brightness / air; variants Still Evening · On a Breeze · Singing.
  • Ticking Clock — a pendulum clock at one beat per second (an even 32 per loop, so the tick/tock alternation survives the seam): tick and tock voiced apart, a wooden case thump under each click, the escapement's tiny secondary contact, and a whisper of room tone. Double lowpass keeps the default sleep-soft. Editors: wood / distance / room tone; variants Distant Hall · Mantel · Close Tick.
• Two new scenes. Curled Up (a warm purr breathing under the exact "Light Rain" variant) and Evening Porch (chimes over night insects and the "Breeze" wind), with their own gradient art.
• All three verified NaN-free, non-clipping, level-matched against heartbeat/night (waveform + spectrogram probes), covered by the existing generator validity / variant-distinctness / determinism tests plus the stereo width (chimes) and centred-mono (purr, clock) assertions. Icon font re-subset for the paw / bell / clock glyphs. 109 unit + 15 e2e, typecheck, lint, build green; console clean.

8.2.0

• Design/a11y polish pass. Three fixes from a full-surface browser audit: the moon no longer hangs over the desktop side panel's pause button (it now offsets past the panel column at ≥1000px) and is pointer-events: none, so scenery can never swallow a tap; the wordmark is now the page's <h1> (screen readers previously landed on "the scenes" with no page title); and the mini player's expand chevron gained a visible keyboard-focus style. Audit also confirmed: token contrast AA-clean, reduced-motion coverage, safe-area insets, and focus-visible styles across all other controls.

8.1.0

• Impact site verification. Added the impact.com site-ownership proof to index.html — both their <head> meta tag and the literal verification text segment in the body (visually hidden) — so it's served in the static HTML at the site root for their crawler.

8.0.0 — Maintenance milestone

A consolidation release closing out a cleanup / documentation / refactor / bug-fix pass. No user-facing feature changes beyond the timer fix; the codebase is leaner, better documented, and better tested.

• Docs. New ARCHITECTURE.md maps the runtime (audio sources, master bus, masking, backgrounding & interruption handling, storage migrations, the PWA self-update flow, and the testing strategy). README project-structure tree refreshed; stray research reports moved to docs/research/.
• Refactor. Sleep-timer logic extracted from App.tsx into a self-contained, documented hooks/useSleepTimer.ts (App.tsx ~66 lines lighter).
• Bug fix. Extending the sleep timer during the final-90s wind-down no longer leaves the mix stuck quiet — the fade now derives from time-remaining unconditionally (windDownFade), covered by a unit test.
• Cleanup. Removed dead exports (generators.ts re-export block, dsp.ts's panMonoInto and SECS export).
• Tests. 100 unit tests + 15 e2e green; added the wind-down fade test.

7.30.1

• Cleanup: removed dead code. Dropped generators.ts's unused individual-generator re-export block (only regenerateSound is the public API), and dsp.ts's unused panMonoInto helper and SECS export. No behaviour change; typecheck/lint/tests/build all green.

7.30.0

• Bug fix: extending the sleep timer mid-wind-down no longer leaves the mix quiet. The wind-down fade only reset to full level when the timer was cleared, never when a running timer was extended back out of the final-90s window — so tapping "+30m" while the mix was fading would leave it stuck at the faded-down level for the rest of the (extended) timer. The fade is now derived from the time remaining unconditionally (windDownFade), so extending restores full level. Covered by a new unit test.

7.29.0

• Refactor: extracted useSleepTimer. The sleep-timer countdown, wind-down fade, expiry, and its handlers moved out of App.tsx into a self-contained, documented hooks/useSleepTimer.ts (App.tsx is ~66 lines lighter). Pure behaviour-preserving move — the timer e2e and full suite stay green.

7.28.0

• Docs & housekeeping. Added ARCHITECTURE.md — a runtime map of the audio engine, master bus, backgrounding/interruption handling, persistence, the PWA self-update flow, and the testing strategy. Refreshed the README's project-structure tree (it was missing graph.ts, layerMeta.ts, several components, and tests/). Moved the two stray background research reports out of the repo root into docs/research/.

7.27.0

• No more faint static after turning every sound off. Toggling a sound off (or pausing) used to re-prime the background keep-alive — the near-silent element that holds the media session — so emptying the mix could leave its faint noise-floor murmur playing with no player bar. Priming is now gated to playback actions only, so when the last layer goes the keep-alive stops with it. (Background audio while a mix is playing is unchanged.)

7.26.0

• Android: another app (or the notification pause) now actually pauses the mix. On Android the media notification and audio-focus are driven by the keep-alive <audio> element, not the Web Audio graph — so when YouTube took focus, or the notification's pause button was tapped, the element paused but the worklets kept playing. The keep-alive's pause is now the cross-platform interruption signal: an unsolicited pause pauses the whole mix, and only an explicit transport action (tap, notification play) resumes it. iOS keeps its AudioContext-interrupted path as well; the two are complementary.
• Crackling is the default fire everywhere. The fire's default character is now Crackling (dry, pop-and-crackle-led over a thin roar) — in the opened editor, a freshly tapped Fire, Reset, the WAV fallback, and presets that don't tune fire. The previous steady character is still available as the Campfire variant.

7.25.0

• Audio interruptions now pause for good — no self-restart. When another app takes audio focus (a phone call, a video, music), the mix pauses and stays paused; it only resumes on a deliberate tap. Previously iOS would mark the AudioContext interrupted and then auto-resume it when the other app finished (and the looping keep-alive element would restart itself), so the sound came back on its own — unwanted in a sleep app. The engine now treats an interruption exactly like a manual pause and pushes any OS auto-resume back down while paused.

7.24.0

• Fire — roar pulled well back everywhere. The continuous roar/body bed is much quieter by default and across every fire scene (Embers, Hearth, Campfire, Bonfire, Wood Stove, Crackling), so the crackle and pop character leads instead of a steady rush. Driven by lowering bodyVol (roar volume) on the default, all variants, and the worklet's own descriptor default; measured roar energy drops ~60%.
• WAV fallbacks now honor every editor control. The procedural fallback loops (used when the live AudioWorklet can't load) were swept so each one responds to its tuning. Two were silently ignoring their params entirely — Fire and Birdsong — so their variants did nothing in fallback mode; both now scale roar/crackle/pop/hiss (fire) and call/trill/peep density, pitch and balance (birdsong) the same way the live sounds do. Rain's space control now nudges the fallback too. A new test asserts every variant of every editable sound renders distinct fallback audio, so this can't silently regress again.

7.23.0

• Thunder — cracks removed entirely, now chip-only. Thunder is now pure rolling rumble: the crack synthesis (click/snap/boom/tear) is gone, leaving the staggered low-frequency roll and its long reverb tail — soft, distant and sleep-safe by design. The fine-tune sliders are removed too; thunder is now selected purely from its scene chips. The chip set was refreshed to four crack-free rumble scenes ordered far → near: Distant Rumble, Rolling Storm (default), Gathering Storm and Heavy Storm.

7.22.0

• "Fan & Rain" scene: fan volume down to 6%. The fan now sits as a faint hush under the rain rather than a prominent layer.

7.21.0

• Thunder — killed the constant fizz, fixed the "taser zap". Two issues: a faint band-limited noise floor ran the whole time on every variant (a persistent fizz), and an overhead strike read as a single electric zap. The always-on floor is removed — between strikes the sound now falls to true silence (the rolling tail decays naturally), which suits sleep. And the close crack was rebuilt: the click and snap are lower and broader (woody crack, not a piercing electric snap), the granular tear is sparser and lower (distinct pops, not a dense buzz), and an immediate deep boom lands right behind the crack and scales with proximity — so an overhead strike is CRACK-boom, not a lone zap. Measured: silent floor (RMS 0), boom 1.5-3× the high-frequency snap energy, NaN-free / non-clipping. 71 unit + 15 e2e, typecheck, lint, build green; console clean.

7.20.0

• Thunder crack made convincing. The close-strike crack was a smooth, buzzy noise burst; it is now built like a real lightning snap: a near-instantaneous broadband click (the percussive leading edge), a bright fast snap right behind it, a short low thump for body/punch, and a dense granular tear — a scatter of tiny clicks that thins out over ~150-300 ms (the electric rip), replacing the old flutter. Measured: a sharp onset (crest ~3-4) and ~8-9 distinct micro-transients in the tear (was 1-4), NaN-free and non-clipping even at close/max-everything. Still procedural, distant storms still skip the crack entirely. 71 unit + 15 e2e, typecheck, lint, build green.

7.19.0

• Thunder reworked for realism (and it no longer opens with a bang). The generator was rebuilt around how real thunder actually sounds: a lightning channel is kilometres long, so its sound arrives over several seconds and rolls. A strike is now a swarm of overlapping low-frequency surges spread across the duration (the rumble climbs, dips and climbs again) rather than one fading hump; a sharp, tearing crack (rapid AM) fires only when the storm is close, so a distant storm is pure roll with no snap; and a long sub-bass swell underpins it. Timing is now two-level and irregular — a drifting storm "activity" plus clustering, so strikes come as a quick flurry then a long lull instead of on a metronome. Crucially for a sleep app, the first strike is held back ~9–21s and softened, and the level eases in slowly — no startle on start. All still fully procedural (no samples). Measured: silent onset (peak ~0.002 in the first 8s), undulating multi-surge envelope, low centroid (~300 Hz distant), NaN-free / non-clipping across every variant. 71 unit + 15 e2e, typecheck, lint, build green; in-browser console clean.

7.18.0

• Thunder — rolling reverb + a crack control (from the thunder research report, kept fully procedural). The report's biggest realism levers were a reverb/echo stage and a crack-sharpness control; both are now synthesised in the worklet, no sample or impulse-response files (the app stays "generated, not borrowed"). A rolling reverb (two damped feedback combs per channel into an allpass diffuser) makes a strike bounce and bloom away; its feedback, damping and wet level ride on distance, so a far storm rolls long and dark while an overhead strike stays tight. A new crack slider sets the high-frequency snap — low for a soft, rounded, distant rumble (the sleep default), high for a sharp close clap. New variants: a soothing Distant Rumble (deep, far, almost no crack, long bloom) alongside Far Off, Rolling Storm, Overhead. Verified NaN-free and non-clipping across every variant incl. max-everything (Node DSP probe); 71 unit + 15 e2e, typecheck, lint, build green; in-browser console clean.

7.17.0

• Fine-tune toggle goes full-width, with a gentle reveal. The "fine-tune sliders" disclosure now stretches the full panel width like the chips and sliders, with its caret pushed to the right edge. Opening it eases the slider groups in (a slight fade + downward settle, lightly staggered) instead of snapping, and the button has a small press response. All stilled under reduced-motion.

7.16.0

• Variant chips line up as a two-column grid. The justified (flex: auto) version grew chips by an equal amount, so rows with different label lengths didn't align. They now use a uniform two-column basis (flex: 1 1 calc(50% - gap)): every full row is a clean 50/50, columns line up across rows, and labels stay on one line (three-up squeezed long names like "Storm in the Trees" onto three lines). A lone trailing chip fills its row.

7.15.0

• Variant chips are justified. The preset chips in the sound editor now grow to fill each wrapped line (flex: 1 1 auto, centred content), so the rows read as justified blocks instead of a ragged right edge — every line spans the full width. The "custom" marker stays its natural size.

7.14.0

• A batch of UI fixes.
  • "your mixes" no longer pops in empty. The section (and its "save this mix" placeholder) only renders once you actually have a saved mix, so starting a single sound no longer shoves the library down. Saving still lives in the player.
  • Save-name focus ring no longer clipped. The "name this mix" field is full-width inside the sheet's scroll area, which clips overflow, so its focus outline was sliced off the sides. The save row is now inset a few px to give the ring room.
  • Presets show the variant they actually play. Loading a preset/scene now syncs each layer's editor values to what's playing, so opening rain's editor during Fan & Rain shows At a Window (not a stale "Steady"). The Fan & Rain preset was aligned to the exact At a Window variant so it reads as a named variant rather than "custom".
  • Fine-tune toggle is one width. Its label no longer swaps between "fine-tune sliders" and "hide sliders" (which changed the button size); it's a single static label, with the rotating caret + accent state carrying open vs closed.

7.13.0

• Hide Stream and Shower for now. Both are pulled from the library via a new explicit HIDDEN_SOUND_IDS set in data.ts (cleaner than marking finished sounds "experimental"; the flag never reveals them). Their generators and editors stay intact, so any saved mix or preset referencing them still plays — delete an id to bring a sound back. Updated the releasableSounds test.

7.12.0

• Cleanup pass. A comprehensive audit (worklet params, runtime console, dead exports, type escapes, both worklets' numerical safety) found the codebase healthy; fixed the genuine inconsistencies it surfaced:
  • Fine-tune focus ring no longer mismatches the button — removed a stale border-radius: var(--r-xs) (6px) override left over from when the control was a text link, so the keyboard outline follows the button's real corner.
  • Now-playing sheet radius used a spacing token (--sp-20) for its corners; it now uses --radius-sheet. .sheet-action and .layer-toggle moved to their semantic radius roles too (no visual change).
  • Removed dead tokens (--radius-detail, --radius-round, --r-xs — all unused) and a dead feature flag (customMixEditor, the editor ships unconditionally). Updated DESIGN.md.
  • Verified both rain and fire worklets are NaN-free and non-clipping across every variant (incl. hiss/metallic at max) via Node DSP probes.

7.11.0

• "Fan & Rain" scene now really is rain on glass. The preset's mood line is "a fan's hush over steady rain on glass," but its rain layer used a generic tuning. It now uses the At a Window variant (glassy taps, roomy, a touch of ring), with the bed trimmed a little so it sits under the fan's hush.

7.10.0

• Hide per-layer mute / solo for now. The M (mute) and S (solo) toggles on each now-playing layer are hidden behind a new layerMuteSolo feature flag (off). The mixer logic stays fully wired up, so flipping the flag back on restores them; the e2e coverage is skipped meanwhile.

7.9.0

• Soft-square design system — phase 2: cards, rename, docs. Scene cards, mix cards and the empty-mix tile move to the 16px --radius-card-lg; sound cards and the inline editor adopt the semantic --radius-card (14px, unchanged value). Renamed .cat-pill → .cat-filter (and .cat-pills → .cat-filters) across the CSS and markup so the category controls read as filter tiles, not capsules, and the naming can't drift back. Documented the whole soft-square radius system (scale + semantic roles table) in DESIGN.md. Focus rings already follow each control's radius (outline / box-shadow respect border-radius), so no change was needed there.

7.8.0

• Soft-square design system — phase 1: controls off the pill. Per the soft-square visual report, controls are now rounded rectangles, not capsules. Replaced the --r-pill (20px) token with a proportional radius scale and semantic aliases (--radius-control-sm 8, --radius-control 10, --radius-button 12, --radius-card 14, --radius-card-lg 16, --radius-sheet 20). Category filters, timer keys, variant chips, the fine-tune toggle, install / cookie / toast / crash actions all move to 8–10px soft squares; the short floating bars (mini-player, toast) drop from a 20px pill to 16px. Circles stay only for genuinely circular objects (play orb, dots, thumbs, slider track). Cards and the rename/focus pass follow in 7.8.x.

7.7.0

• Fine-tune control reads as a button. The sound-editor disclosure that reveals the per-parameter sliders was bare muted text with a small chevron, so it was easy to miss and unclear what it did. It now uses the shared pill vocabulary (bordered, tappable), leads with a tune (sliders) icon, states the action ("fine-tune sliders" → "hide sliders"), and takes the category-accent "open" state of a real toggle.

7.6.0

• Variant chips gain crafted marks. Each preset chip in the sound editor now leads with a small 16px mark (new src/lib/variantIcons.tsx, drawn in currentColor so it inherits the chip's idle/active colour). Intensity sounds show 1/2/3 ascending bars; rain shows fine drizzle → 1/2/3 drops, then the three surface scenes (peaked roof, glazed window, corrugated tin) grouped last; fire shows per-character flames (embers, hearth arch, campfire, bonfire, wood-stove box, crackling sparks). The variant name stays the code identity and the visible, accessible label (marks are aria-hidden); a couple of lists were reordered so the bars read low→high. Verified in-browser (mobile screenshots) and via 71 unit + 16 e2e, typecheck, lint, build.

7.5.0

• Rain — open-air drops become a super-muted soft patter. The open-air variants (Light Rain, Drizzle, Steady, Downpour) keep their drops, but as a soft, dull patter sitting under the wash — rain on soft ground or a tent canvas — rather than discrete taps even with the bed. The dropGain floor is lowered (0.8 → 0.5) so low drops genuinely recedes, and the open-air variants drop to roughly half their previous hit level (e.g. Steady dropGain 1.34 → 0.71). At metallic 0 the hits are already dull/dark (the de-tinning), so muted reads as soft, not tinny. The surface variants (On a Roof, At a Window, Tin Roof) raise drops to compensate exactly, so their crisp hits are unchanged (e.g. Tin Roof dropGain stays 2.36). WAV fallback floor matched. Verified: 71 unit + 16 e2e, typecheck, lint, build green.

7.4.0

• Rain — Tin Roof no longer rings like wind chimes. At high metallic the solid drops each fired a ~90 ms pure sine at a random 2.4–5 kHz pitch, so a tin roof played a chord of different notes. Real heavy rain on tin is bright broadband drumming; a tuned ring only suits sparse drops on glass. The tonal ring now peaks at moderate metallic (≈ At a Window) and fades out as metallic climbs — at full Tin Roof the rings drop from ~42% of drops to ~7%, 6× quieter and 3× shorter, with the broadband brightness carrying the metal instead. The ramp is anchored so At a Window (metallic 0.30) is left exactly as it was. Tin Roof stays bright (measured centroid ≈3380 Hz, above Window ≈2530 and the ≈1880 default). Verified: 71 unit + 16 e2e, typecheck, lint, build green.

7.3.0

• Rain — the default no longer sounds like rain on tin. The 7.2.0 metallic control was opt-in but never actually darkened the baseline, so every preset still rang in the bright glass/metal register. This reworks the default to be genuinely soft rain on the ground and makes brightness ride entirely on the metallic macro: drop centres sit low and broad (≈430–1000 Hz, not 1.5–3.4 kHz) and only rise with metallic; the water plips lose their tonal high-Q ring; the colored 4 kHz/8 kHz bed bands sit well down; and a clean 2-pole master lowpass (dark by default, opening with metallic) caps the whole shower. Measured spectral centroid of the default fell from ≈4440 Hz to ≈1870 Hz (energy above 2.5 kHz 55% → 26%), with a clean brightness gradient up through At a Window (≈2510 Hz) to Tin Roof (≈3390 Hz). The WAV fallback is darkened to match. A worklet-DSP probe (run in Node) caught a runtime ReferenceError the unit tests can't reach, since the worklet isn't unit-tested. Verified: 71 unit
  • 16 e2e, typecheck, lint, build green; peaks well below clipping, no NaN.

7.2.0

• Rain — opt-in metallic surface + clearer surface variants. The rain worklet gains a metallic control: at 0 the default stays natural and de-tinned (rain on soft ground), but raising it brings the solid drops' ring back on purpose — brighter (2–5 kHz band), louder, longer and more frequent — so a real tin-roof surface can ring without that edge leaking into every preset. Two new variants use it to pull the surfaces apart: Tin Roof (sharp ringing pings) and a glassier At a Window (a touch of ring, roomier), while On a Roof is now wooden and muffled with almost no ring. A Light Rain variant adds the very sparse, intermittent end the intensity range was missing. Verified: 71 unit + 16 e2e, typecheck, lint, build green.

7.1.0

• Fire — flame hiss band + distinct presets. The fire worklet gains a dedicated high-band hiss (one-pole high-pass → gentle low-pass), gated by flame energy and the roar envelope so it breathes with the fire instead of sitting as flat air. A new hiss control exposes it. This unlocks textures the old roar-plus-crackle model couldn't reach — most of all a Wood Stove preset (dark low rumble, prominent escaping-air hiss, only a few distant pops). Presets are retuned to actually diverge rather than differ by volume: Embers smoulders (sparse soft crackle), Hearth is a steady close fireplace, Bonfire lets the roar dominate with frequent loud pops and little hiss, Crackling stays dry-kindling rapid. Verified: 71 unit + 16 e2e, typecheck, lint, build green.

7.0.0

• Rain & fire realism — opening the 7.x line. A focused pass on the two generators the research report singled out: rain that reads as rain on soft ground rather than on tin, with intensities and surface variants that actually diverge; and fire whose presets are distinct textures (stove, embers, hearth, campfire, bonfire) rather than one bonfire at different volumes. This major bump opens the line; the work lands across the 7.x minors that follow. No behaviour change in this entry itself — it sets the baseline.

6.5.0

• Audio engine — sprint 6: final check + fix. Closes the 6.x audio pass. Documented the unified mix graph, per-layer EQ, two-dimensional masking, mute/solo, and sleep-safe in DESIGN.md. Fixed a mute/fade race (a fade-in started just before a mute could ramp the volume back up) by reading mute/solo from refs so live fade timers see current state. Verified end to end: 71 unit + 16 e2e (incl. audio-flow, mute-silences, sleep-safe), typecheck, lint, build all green.
  • Note for device testing: WAV layers route through the graph via MediaElementAudioSourceNode, which is verified on Chromium but can be silent on some iOS Safari versions; the code falls back to direct output if routing is refused. Please confirm background / lock-screen playback on an iPhone.

6.4.0

• Audio engine — sprint 5: mute / solo + sleep-safe toggle. Every layer in the now-playing controls gains compact M (mute) and S (solo) toggles — silence a layer or isolate it without removing it. A sleep-safe toggle in the master section (on by default, persisted) enforces the calmer DSP policy: a darker master high-shelf and the spectral slotting of stacked broadband beds; turn it off for a brighter, more cinematic balance. Also fixed a fade race so muting a just-started layer takes effect immediately.

6.3.0

• Audio engine — sprint 4: sleep-safe defaults. Retuned the default character of most layers toward calmer, sleep-first listening (from the research report's sleep-safe table): a fainter, softer rain bed with fewer drops; quieter, darker fire crackle; wind whistle way down; sparser, more distant thunder; less twig detail in the forest; lower ocean foam and stream sparkle; darker white noise; warmer pink/brown noise; quieter night insects. WAV loops now render from their editor defs (single source of truth for a default), so a default lives in one place. Named variants and saved mixes are unaffected; every value is still editable.

6.2.0

• Audio engine — sprint 3: per-layer EQ + spectral masking. Every layer now runs through its own bus (lowpass → high-shelf → trim) into the master bus. Masking is no longer just a level trim: when more than two broadband/water beds stack, the non-accent ones move out of each other's way — a darker top, a high-shelf cut, and a small extra trim — so the busiest layer keeps the spectrum and the rest recede. Solo sounds and small mixes stay transparent; accents keep their top. (Sharpness drives listening fatigue more than level.)

6.1.0

• Audio engine — sprint 2: WAV loops join the bus. The rendered loop layers now feed the shared graph too, via MediaElementAudioSourceNode → master bus. The proven HTMLAudioElement crossfade still drives playback (so background / lock-screen behaviour is preserved), but its output is processed with everything else — the whole mix is finally gain-staged together. Falls back to direct element output if a platform won't allow the routing. Added a master analyser/peak meter (verified audio flows through the bus in-browser).

6.0.0

• Audio engine — sprint 1: shared master bus. Introduced one Web Audio graph with a master bus (src/audio/graph.ts): a gentle glue compressor, a quiet high-shelf to take the edge off the top, and a fast safety limiter against clipping. The live worklet sounds (rain, fire, thunder, windy forest, birdsong) — the most transient, harshness-prone layers — now play through it instead of straight to the speakers. The chain is ~loudness-neutral on a single sound; it only works when layers stack. (Start of an audio-architecture pass; major bump to open the 6.x line. WAV loops join the bus next.)

5.5.1

• Fix: save name field too small on phone. The save row now takes its own full-width line below the actions, with the name field on its own line above a full-width save button, instead of a cramped inline field.
• Fix: fire crackle sounded bubbly again. Stabilising the worklet pan (5.x audio work) removed an old per-channel smoothing that had been masking the crackle's resonant ring; on clean stereo those tonal pings read as water bubbles. The crackle bandpass Q is dropped to a broad, dry band and blended with a raw transient, so each crackle snaps rather than plinks.

5.5.0

• UX redesign — sprint 6: review, refactor, docs. Closes the 5.x UI/UX pass. Verified the whole flow across mobile and desktop, documented the new layout and components in DESIGN.md, and locked behaviour with e2e coverage for the desktop side panel, stop/undo, save-from-player, and delete/undo. No new surface; consolidation of sprints 1–5.

5.4.0

• UX redesign — sprint 5: accessibility + contrast. All three text tiers now clear WCAG AA for normal text: --text-dim was at 3.4:1 (below the 4.5:1 floor for the meta-count labels that use it) and is raised to 4.6:1, with --text-secondary lifted to 5.5:1 so the tiers stay visibly ordered. The small card ✕ controls and the toast dismiss gained wider invisible touch halos (toward the 44px guideline) for one-handed use in the dark.

5.3.0

• UX redesign — sprint 4: scene clarity + microcopy. Scene cards now carry a quiet layer count (e.g. "2 layers"), so they read as real presets — the icons already show which sounds, the count shows how many — matching the saved-mix cards. On wide screens the scenes wrap into a grid (every scene visible, no horizontal-scroll trick) while mobile keeps its peeking shelf. The install row copy is clearer: "install for a faster start and offline nights."

5.2.0

• UX redesign — sprint 3: desktop/tablet split layout. Wide screens (≥1000px) are no longer a phone-width column stranded in the middle. The shell becomes a two-column grid: browse on the left, a persistent control panel on the right that shapes the mix in place (layers, master, sleep timer, drift, save) instead of through the slide-up sheet. The now-playing controls were extracted into a shared MixControls component, so the mobile sheet and the desktop panel can never fall out of sync. Mobile is untouched.

5.1.0

• UX redesign — sprint 2: save discoverability + undo. Saving a mix no longer hides a layer deep in the now-playing sheet: a bookmark sits in the mini player and opens straight into the name field, and while a mix is playing with nothing saved yet, the "your mixes" section shows a quiet card that teaches the feature and doubles as the save entry. Destructive actions are now forgiving in the dark: "clear all" becomes "stop mix" and both stopping a mix and deleting a saved one leave a five-second undo snackbar.

5.0.0

• UX redesign — sprint 1: prompt timing. First load now stays clear: the storage notice and the install row are held back until the first sound has actually played, and they never appear together. The storage notice shows first (once, then remembered); the install row only surfaces once that's acknowledged or was on a prior visit. Nothing competes with the path to sound. (Start of a UI/UX pass; major bump to open the 5.x line.)

4.5.2

• Fix: wind sounded like two large fans. Wind's whoosh is a low/mid-band source (rolled off ~1–1.4 kHz), so the default 800 Hz decorrelation crossover was spreading the body itself — the gust split into two sources left and right. Wind now uses a higher crossover (1.6 kHz) so the whole whoosh stays centred; its width comes from the panned edge-tone whistles instead.

4.5.1

• Fix: decorrelated beds sounded like two separate sources. Pink/white noise and the other broadband beds were widened by shifting the whole band in opposite directions per channel, which drives the interaural correlation toward zero on steady noise — the ear hears that as two uncorrelated sources on either side rather than one wide source. decorrelateMono now keeps the low/mid band shared (a fused mono centre, where the binaural system takes its localisation cue) and only spreads the highs (bass-mono / treble-wide). Spacious but coherent, still no comb filtering. (Fan, brown noise, heartbeat and the underwater body remain intentionally centred.)

4.5.0

• Stereo overhaul — sprint 6: hardening & docs. Locked the stereo and masking work behind regression tests: generators are asserted to render two-channel loops, broad beds to be genuinely decorrelated (L/R correlation below threshold) while compact sources (brown noise, fan, heartbeat) stay centred, and layeringTrim to apply the expected same-group cuts. Documented the stereo/masking model in DESIGN.md and the README. No audible change — this sprint makes the previous five safe to build on.

4.4.0

• Stereo overhaul — sprint 5: worklet spatial polish. Fire and Birdsong were jittering their pan every sample, which reads as a constant stereo "fizz" rather than a located source. Both now hold a pan position and ease between positions: the fire sways gently as one source, and each bird call/trill jolts the placement so calls arrive from new directions across the field. (Measured: fire's abrupt pan jumps dropped from hundreds to ~0 per 1000 samples.)

4.3.0

• Stereo overhaul — sprint 4: masking-aware layering. Added layer roles (src/audio/layerMeta.ts): every sound is a bed, motion, or accent in a spectral mask group (broad / water / low / detail). When layers in the same group stack, the mixer now applies a gentle automatic level trim — beds duck a little for each broadband neighbour, a crowd of same-group motion layers ducks the extras, accents are left alone — so stacking (Fan + White Noise + Rain, say) stays clear instead of fogging into one lump, without hand-tuning every preset. (Graph-backed WAV playback for per-layer panning is deferred to a session where the audio can be checked by ear.)

4.2.0

• Fixed the "jet engine" comb artefact in the stereo beds (4.0–4.1). The widener was blending each channel with a short delayed copy of itself, which comb-filters noise into a flanged whoosh on both sides. It now decorrelates by pure opposite time-shifts, so each channel keeps a flat spectrum — wide and diffuse, no flanging.
• Stereo overhaul — sprint 3: urban & deep beds. Train rolls around you (decorrelated floor + spread joint clatter; underfloor thumps stay centred), Airplane widens its boundary-layer "blanket" while the engine body stays centred, Underwater scatters its bubbles across the field over a centred deep body, and Night Insects place their chirp bands around you. Brown noise, Fan, Heartbeat and the underwater body stay mono on purpose (compact or non-directional sources). All WAV sounds are now stereo-aware.

4.1.0

• Stereo overhaul — sprint 2: water, air, and noise beds. Converted the broad ambience layers to stereo: Stream (body and bright ripples decorrelated separately, so it glitters wide), Wind (decorrelated gust bed with each edge-tone whistle placed across the field), Shower (wide spray + a roomier reflection spread), and the White / Pink / Brown noise beds (gentle width; brown stays near-centre since bass isn't directional). Tests assert each is genuinely not mono.

4.0.0

• Stereo audio overhaul — sprint 1: foundation. First step of moving the procedural sounds from a flat mono wash to a wide, moving stereo image. Added stereo rendering to dsp.ts (a 2-channel WAV writer, genStereo, equal-power panMonoInto, and decorrelateMono for widening beds), and converted Ocean to it: a decorrelated undertow bed with surf that crests and pans across the image instead of pulsing dead-centre. Hardened CrossfadeAudio.destroy() to revoke blob URLs (no leaks over long retuning sessions). New tests cover stereo seam continuity and L/R decorrelation.
• Fixed the rain editor's "fine-tune" control showing literal "expand_more" text — that glyph isn't in the subsetted icon font; switched to the bundled keyboard_arrow_down.

3.9.0

• Sound "variants" + progressive disclosure. Opening a sound now leads with a row of named character presets (chips) instead of a wall of sliders — e.g. rain → Drizzle · Steady · Downpour · On a Roof · At a Window, fire → Campfire · Embers · Hearth · Bonfire · Crackling. Every editable sound gets 3–5 evocative variants. The full sliders moved behind a "fine-tune" disclosure, collapsed by default, so the three depths are: card (toggle + volume) → variant chips → sliders. The default character is unchanged for every sound. Picking a chip applies its values live; nudging a slider shows a quiet "custom" marker.

3.8.2

• Decluttered the rain editor (10 sliders → 7). Folded overlapping controls that shape one thing into single macro sliders: the bed's tone now follows surface (a soft surface dulls the whole shower), stereo width rolls into space (room + width together), and the slow swell plus the drop clustering combine into movement (steady → alive). Renamed the rest for plainness (bed level → background, drop hits → drops). The default sound is unchanged and the Fan & Rain bed control is untouched. Reviewed the other sounds' slider names too; they were already concrete, so they're left as-is.

3.8.1

• Rain follow-ups: softened the drop impacts so they read as taps rather than a tin-roof ping (lower centre frequency, gentler filter Q, rarer/fainter tonal ring; the WAV fallback's ping was thinned to match) and lowered the default drop level. Fan & Rain's bed is back, just a bit fainter than normal (0.7).

3.8.0

• Per-scene sound tuning. Built-in scenes (and the resume-your-night session) can now carry editor slider overrides per layer, not just a level — so a scene can shape a sound's character. Applied on load and reset when a sound plays on its own, so the standalone sound keeps its normal settings.
• Fan & Rain fix. The rain layer there brought its own broadband bed, which doubled up with the fan's hush. The scene now drops the rain bed right down (new bed level control at 0.18) so the fan carries the wash and the rain just adds drops over it. Standalone rain is unchanged.
• Rain bed controls. Two new sliders in the rain editor: bed level (how much of the continuous curtain sits under the drops) and bed tone (dark and soft → open and airy). The bed bands were also broadened (lower filter Q) and the top trimmed, taking the metallic "tin-roof" edge off the curtain.
• Lengthened each scene's mood line by a couple of words, and clamped the line to two so a longer description tapers cleanly on the fixed-height card.

3.7.4

• Polish pass — lighter first paint. Code-split the heavy, post-startup pieces out of the initial bundle: the procedural WAV-generator DSP (src/audio/generators.ts, ~6.8 kB gz) now loads on a sound's first play, and the now-playing sheet, drift mode, and sound editor load on first open (all idle-prefetched after first paint so opening stays instant). The WAV source path (data.ts → audio/sources.ts) is now async to support this. Initial app JS dropped ~72 → ~63 kB gz. The service worker auto-precaches the new chunks, so offline is unaffected (verified by the e2e suite).
• Fixed: under prefers-reduced-motion, rotating/resizing the window cleared the NightSky canvas and left it blank (no animation loop to repaint) — it now redraws the still frame on resize.
• Refactors: extracted the shared "start a source with fade-in" logic in useAudioMixer (deduped across toggle / resume-all / restore-mix); memoized the library/visible-sound lists in App so they don't churn effects each render; corrected the storage/keys.ts doc comment.

3.7.3

• Rain: softer by default. Lowered the default drop hits (0.70 → 0.35) so the surface impacts sit under the bed (gentle, not harsh); the slider still reaches fully forward.

3.7.2

• Fixed: the mini player covered the footer (privacy/version) intermittently. The clearance was a hardcoded gap; the player's real height varies with safe-area insets, font scaling, and subtitle wrap. It now measures the live player (ResizeObserver → --mini-player-h) so the footer always clears it.

3.7.1

• The scenes shelf "there's more this way" wink now plays on every load, not just the first (the one-time localStorage guard was removed).

3.7.0

• Rain rework. Moved the rain worklet from a single filtered-noise bed + uniform-Poisson ticks toward a richer texture, keeping the same browser stack: a multi-band, decorrelated bed whose sub-bands drift on slow random walks (so it breathes instead of hissing); a two-level scheduler (steady background drizzle + occasional clustered bursts) so the rain clumps and lulls; surface- aware drop atoms (hard → bright/snappy, foliage/wet → darker/softer); and a short multi-tap early-reflection bus for placement. Added four character controls to the editor — drop hits, patter, space, stereo width — alongside the existing intensity / heaviness / surface / swell.

2.11.3

• Renamed the "tonight" section heading to "the scenes" (meta line now "curated mixes"), matching the "the library" naming pattern.

2.11.2

• Fixed the odd line under the presets shelf. The scene/mix rows are horizontal scroll containers, which clip vertically — so the cards' soft drop shadows were being sliced into a hard dark band at the shelf's bottom edge (background looked different below vs above). Removed the outer drop shadows on shelf cards; the border + inset top hairline carry the depth.

2.11.1

• Fixed: couldn't reinstall after uninstalling. Installing wrote a permanent "dismissed" flag to localStorage, which outlives the app — so after an uninstall the in-app install prompt never returned. Installing no longer writes the flag (Chrome stops firing beforeinstallprompt while installed anyway, and clears it instead), dismissal now hides the row for 14 days rather than forever, and the old legacy flag is treated as expired so existing devices recover automatically.

2.11.0

• App-icon shortcuts + scene deep links. Long-press the installed app icon (Android; right-click on desktop) for Rainfall / Fireside / Deep Rest shortcuts. They use ?scene=<id> deep links, which the app now resolves on launch (autoplay-blocked launches load the mix paused, one tap to play). Manifest also gains id, lang, dir, and display_override.

2.10.0

• Offline support. Added a service worker (public/sw.js, registered in production): the app shell, hashed assets, icons, worklets, and fonts are cached so drift away opens and plays with no signal — sounds are generated on-device, so offline is fully functional. Navigations stay network-first, so deploys are picked up on the next online visit.

2.9.0

• iOS background playback + best-possible media notification. New src/lib/backgroundAudio.ts: declares navigator.audioSession.type = 'playback' (iOS 16.4+) so audio keeps playing when the tab is backgrounded or the screen locks, and runs a looping silent <audio> keep-alive (started inside the play gesture, synced to play/pause) so the lock-screen / notification player reliably appears even for worklet-generated sounds, on both iPhone and Android.
• Polished webapp head. Brand-blue browser tab/address-bar (theme-color #7ba7e8, with a dark-scheme variant), color-scheme: dark, Open Graph + Twitter cards (icon artwork), canonical URL, application-name, mobile-web-app-capable, and format-detection off for phone numbers.

2.8.3

• Fixed the scene cards going solid ("something pops in behind"). The culprit was backdrop-filter on the preset cards: over the near-black sky the blur had nothing useful to frost, so it composited a near-solid dark layer that popped in behind the translucent gradient. Removed backdrop-filter from the scene and mix cards (and reverted 2.8.2's frosted-white film). The cards are now just their translucent colour gradient over the night sky — colour + transparency, no pop. The mini player, sheet, and storage notice keep their blur since they sit over real content.

2.8.2

• Fixed the scene cards looking "transparent on fade-in, then solid at rest." They were a dark colour gradient with no frosted base, so over the near-black sky they read as solid (the only translucency was the entrance fade). Gave them a faint frosted-white film under the colour tint — the same trick the mix cards use — by moving the gradient to background-image and adding background-color: rgba(255,255,255,0.06). They now read as a lighter glass pane, tinted by the scene colour, at rest.

2.8.1

• Preset cards: the gradients were still too dense to see through (a regex had also nudged some colour radials up to 0.30). Rewrote each scene's art with a genuinely faint tint — colour radials ~0.16–0.24, near-clear base (~0.08/0.03) — so the frosted night sky shows through the cards. The bottom scrim still carries the name.

2.8.0

• Preset cards are now genuinely transparent. v2.7's glass treatment kept the gradient base at ~0.55 alpha, which over the near-black background still read as solid. Dropped the base gradient stops to 0.30 / 0.18 alpha so the night sky shows through the cards (the colour tint + blur + sheen remain); strengthened the bottom text scrim so names stay legible.

2.7.0

• UI review & fix-up — tinted-glass preset cards. The scene and mix cards keep their colour gradients but now read as proper tinted glass: a backdrop blur + saturate frosts the night sky behind, a light-catching top hairline (inset 0 1px 0) and a top-left sheen sell the glass, and a soft drop shadow lifts them off the page (instead of flat gradient tiles).
• Unified glass material across all five glass surfaces — scene cards, mix cards, mini player, now-playing sheet, and the storage notice now share the same blur + saturate + highlight recipe, so they feel like one material.
• Removed two orphaned design tokens (--surface-active, --shadow-card); DESIGN.md updated. Revert point for this pass: branch backup/pre-ui-review-v2.6.0.

2.6.0

• Storage notice ("cookie" popup). drift away sets no tracking/advertising cookies, so instead of a misleading banner it shows an honest one-time note: mixes and settings are kept on-device (localStorage), nothing leaves it, no tracking/ads — with a privacy link and a "Got it" button, styled to match the playback bar's glass. Remembered after dismissal.
• Train rebalanced. A broadband pink-noise "carriage" wash (mixed at 0.33) was overwhelming the wheel-clatter rhythm (only 0.10). The carriage bed is now darker and much quieter (0.11) and the clatter leads (0.24), so the train reads as a train.

2.5.1

• Reordered the scene shelf to lead with the strongest presets: Rainfall, Fireside, Deep Rest, then the rest.

2.5.0

• Renamed the app to "drift away" — wordmark, page title, PWA manifest (name + short name), apple-web-app title, media-session metadata, and the privacy page.
• Fixed the slider "detached ball" look. The thumb's soft glow over a thin 4px track read as a floating ball beside a rectangle. The track is now 6px, the thumb is grounded with a tight drop shadow (no glow), and ::-moz-range-track is defined for consistent cross-browser centering — the thumb now reads as a knob sitting on the bar.
• Made the playback bar stand out. It was too translucent to register: now a more opaque, defined glass bar with a stronger lift shadow and a defining edge, and the play control is a solid bright accent disc (dark glyph) as the one high-contrast focal point.
• Deep pass on the library. The category filter pill now glows in its own family hue when active (Fire ember, Air cyan, …), tying the row to the cards; active sound cards are tinted with their category color instead of a generic blue-white; sound names are a step larger/clearer; counts and icons refined.

2.4.0

• Removed the Café sound and its "Café in the Rain" scene (library is now 18 sounds, 7 scenes). The Underwater scene is now just underwater (dropped the night drone).

2.3.2

• Trimmed scene layers per feedback: Rainfall is just rain (no fan), Windy Forest is just wind + forest (no birdsong), Fireside is just fire + a 5% night drone (no wind).

2.3.0

• Reworked the eight curated scenes with proper level balance (not everything at 0.5). Each is mixed so the focal layer leads and broad beds sit underneath: Rainfall (rain 0.60 · fan 0.18), Distant Storm (thunder 0.62 · rain 0.40 · wind 0.16), Windy Forest (forest 0.60 · wind 0.30 · birdsong 0.12), Fireside (fire 0.62 · wind 0.20 · night 0.05), Ocean Night (ocean 0.55 · wind 0.18 · night 0.08), Deep Rest (brown noise 0.50 · heartbeat 0.24 · night 0.12), Underwater (underwater 0.60 · night 0.20), and Café in the Rain (café 0.55 · rain 0.30). Scene art and mood lines re-synced to the new set.

2.2.0

• Library curation. Removed Waterfall, Tent Rain, Tin Roof Rain, Frogs, Dryer, and Deep Space. Night now plays the Deep Space synthesis (the deep, drifting void was the keeper), so its slider set is now Void / Shimmer / Drift. The library is now 19 sounds.
• Campfire Night reworked: the Night drone at 2% under Fire (was Fire + Forest). The Underwater Cave scene and the Rainforest scene were updated to drop the removed layers (Underwater + Night; Rain + Windy Forest + Birdsong). Dead generators and editor entries removed.

2.1.2

• Removed static noise floors. Thunder no longer carries a constant pink "air" bed — it now lives over true silence between events. Underwater's hissy midrange pink "current" wash is gone; its movement now comes from a slow swell of the deep rumble, with a darker final low-pass so no high-frequency static survives.
• Translucent scene & mix cards. The curated-scene cards bottomed out in opaque colours; their gradient art is now translucent over a backdrop blur (and mix cards gained the same glass), so the night sky shows through and they sit in the same material as the mini player and now-playing sheet.

2.1.1

• Rain de-laser tuning. Raindrop hits read as sci-fi "pew" sounds because the water family used a descending sine glide (a literal laser zap) and the other drops used high-Q bandpasses that ring tonally. Fixed by removing the glide entirely, dropping filter Q across all drop families so each is a broadband tick rather than a tuned beep, and making resonant rings faint, short, and rare. Water plips are now low, soft, and infrequent.

2.1.0

• Real-time synthesis for Rain, Thunder, and Windy Forest (sound generation overhaul, fully in-browser via AudioWorklet). These three were the least recognisable because they were stationary noise loops; environmental sounds are identified from temporal events, not static spectra. Each is now generated live off the main thread (public/worklets/), with the previous WAV kept as an automatic fallback if the worklet can't load:
  • Rain (rain.worklet.js): an airy filtered-noise bed under three drop families — solid hits (with occasional modal rings), leaf hits, and water chirps — scheduled as independent Poisson processes across near/mid/far lanes with slow density drift. You hear impacts first, bed second.
  • Thunder (thunder.worklet.js): sparse storm events over silence, each a sequence of 1–4 staggered claps → a downward-sweeping rumble → an undulating afterimage → a sub-bass deepener, all from one seed (62% of strikes are 1–2 claps, per the model).
  • Windy Forest (windyforest.worklet.js, renamed from "Forest"): four wind-speed-driven canopy bands with leaf rustle emitted as a child of the gust field (rustle swells with the whoosh), plus rare branch creaks/whistles on stronger gusts.
  • The 3-slider editors for these now drive the worklet live (k-rate params) instead of regenerating a WAV; the generic worklet source + WAV-fallback path is unified in useAudioMixer.ts (fire reuses it too). Saved mixes are unaffected (sound ids unchanged).

2.0.0

• Full product overhaul (revert point: branch backup/pre-overhaul-v1.2.0 / tag v1.2.0-pre-overhaul). The in-browser sound generation engine is untouched; everything above it was rebuilt around a browse-first, player-persistent architecture:
  • Scenes: the eight built-in presets become curated gradient-art cards in a snap-scrolling shelf, each with a mood line and CSS-generated art (src/lib/scenes.ts). Tap to play instantly; tap the playing scene to pause. Hand-editing the mix clears the scene badge.
  • Mini player (src/components/MiniPlayer.tsx): a floating glass pill at the bottom whenever a mix is active: play/pause with the sleep-timer ring, serif mix title, countdown/layer count, one tap into the sheet.
  • Now-playing sheet (src/components/NowPlayingSheet.tsx): the mix's control room. Per-layer category-colored sliders with remove, master volume, the sleep timer (with "ends ~11:42 pm"), and doorways to drift mode and saving the mix. Replaces the old master bar entirely.
  • Your mixes: saved presets as warm serif cards with layer counts, replacing the chip row. Saving lives in the sheet.
  • Typography: Cormorant italic now carries the brand moments: greeting ("good evening"), section headings (tonight / your mixes / the library), scene and mix names, sheet title.
  • Atmosphere: a slow aurora layer (indigo/violet/teal, 90s transform loop) joins the starfield; stilled under reduced motion.
  • Details: time-of-day greeting, spacebar play/pause, footer simplified (rest well · privacy · version inline; privacy moved off fixed positioning since the mini player now owns the bottom edge), formatCountdown deduplicated into src/lib/time.ts.

1.2.0

• Reusable install prompt: Chrome only shows its native install banner once, then suppresses it. Added an early beforeinstallprompt capture (index.html) plus an in-app install row (src/components/InstallPrompt.tsx) that appears under the header whenever the app is installable and triggers the native prompt on demand. It hides once installed or dismissed, never shows when already running standalone, and falls back to a short "Add to Home Screen" hint on iOS Safari (which has no install event).

1.1.1

• Pinned the footer "privacy" link to the very bottom of the screen (fixed, centered) instead of trailing the "rest well" footer line.
• Made the privacy page typography match the app: loads Cormorant + Inter, with the Cormorant italic wordmark replacing the Georgia serif fallback.

1.1.0

• New app icon: redrew the icon as a gold crescent with a soft halo over a radial night sky (public/icon.svg), regenerated to all PNG sizes (192/512/1024 + 180 apple-touch) and a refined small-size favicon.svg. Manifest icons split into proper any + maskable entries.
• Better media-session player: the notification/lock-screen player now leads with the mix names as the title (with "drift" as the artist line), ships multiple artwork sizes, and uses a redrawn crescent-night artwork that matches the app icon so Android tints the notification on-brand.
• Android system bars: theme_color and background_color set to the true background #080c14 so an installed PWA themes the status bar (and, on recent Android, the navigation bar) to match. (The bottom nav bar can only be fully controlled from a native/TWA wrapper, not from the web layer.)
• Sound editor: removed the developer "config values" export section (JSON + copy button) entirely, along with its now-orphaned styles.
• Privacy & terms: added a hosted policy page (public/privacy.html, linked from the footer) and a repo mirror (PRIVACY.md), covering the no-data-collection model, local-only storage, the fully self-hosted (no third-party) and offline design, and a not-a-medical-device disclaimer.

1.0.1

• Replaced the per-card icon watermark with a soft corner color wash. The large faint category glyph (and its per-family motion) read as decoration and crowded each card alongside its real sound icon. Each card now carries only a gentle radial glow of its category hue in the top-right corner (.sound-card::after): invisible at rest, a hint on hover, gentle when active. The per-category color identity stays; the glyph is gone.

1.0.0

• Onboarding, accessibility, and ship-grade hardening (Stage 6, completing the design roadmap):
  • First-run whisper: a single Cormorant line above the grid ("tap a sound to begin · layer as many as you like"), shown until the first sound is ever toggled, then remembered via localStorage.
  • Accessibility pass: aria-labels on every slider (master, per-sound, editor params), aria-pressed on category pills and timer chips, decorative elements (indicator, watermark, ring, version stamp) hidden from assistive tech; drift mode already carries dialog semantics with focus management.
  • Expanded hit targets: small chips keep their calm visual size but gain invisible touch halos approaching the 44px guideline; touch-action: manipulation removes double-tap zoom delay; tap highlight removed. Pinch zoom re-enabled (dropped maximum-scale=1 from the viewport).
  • PWA/meta fixes: manifest.json icon paths and start_url/scope made relative (they 404'd under the /SleepSounds/ GitHub Pages base), manifest and theme-color aligned to the true background #080c14, media-session artwork fallback made base-aware, and a meta description added.

0.6.0

• Per-sound character + tactile micro-interactions (Stage 5 of the design roadmap):
  • Category accents: each sound family now carries its own muted hue when active (Fire ember, Air pale cyan, Earth sage, Noise dusty violet, Urban steel, Wildlife moss, Cozy warm gold; Water keeps the signature blue). Applied to the active card's border, glow, icon, indicator dot, equaliser, volume slider, and editor handle. All hues ≥7:1 contrast on the background.
  • Element watermark returns: reinstated the per-card category watermark (it had become orphaned CSS after an earlier refactor) and gave each family motion that fits its nature while playing: water bobs, fire flickers, air/wildlife sway, cozy breathes; earth, noise, and urban hold still.
  • Tactile controls: slider tracks 2→4px and thumbs 10→14px with a soft glow and press scale; cards compress slightly on press; new preset chips ease in when saved; long preset names now ellipsize.
  • Haptics: tiny vibration confirmations (8–12ms, best-effort, no-op on iOS Safari) on sound toggle, play/pause, timer set, and preset save.

0.5.0

• Drift mode + wind-down timer (Stage 4 of the design roadmap):
  • New fullscreen drift mode (src/components/DriftMode.tsx), entered via the moon button in the master bar once sounds are active: the mixer chrome fades away, the living sky shows through, and a large Cormorant clock, the mix names, the countdown, and a breathing play orb remain. Controls quiet down after 5 still seconds (the clock stays); any touch wakes them. Esc or the close button exits; exiting is automatic if the mix empties.
  • Requests a screen wake lock while drift mode is open (best-effort, re-acquired on tab return) so the bedside display stays softly lit.
  • Timer redesign: the −30/+15/+30/✕ increment chips are replaced with absolute durations (15m · 30m · 1h · 90m); tap the active chip to cancel. A warm progress ring around the play button empties as the timer runs.
  • Gentle wind-down fade: over the timer's final 90 seconds the mix eases out (playback gain only, via a new masterFade multiplier in useAudioMixer) so the stop never jolts. Sound-generation algorithms untouched.

0.4.0

• Living night sky (Stage 3 of the design roadmap): replaced the static CSS starfield with a procedural canvas scene (src/components/NightSky.tsx):
  • Seeded star placement, so drift's constellation is identical every visit
  • Stars twinkle on individual phases; the sky brightens gently while the mix plays and settles when idle
  • During the last five minutes of the sleep timer, the sky gradually dims with the wind-down
  • A rare meteor crosses while playing (never when idle or during deep wind-down)
  • 30fps render cap, DPR capped at 2, pauses on hidden tabs, and draws a single static frame under prefers-reduced-motion

0.3.0

• Design-system foundation + typographic refinement (UI/look-and-feel; no audio or behaviour changes):
  • Added DESIGN.md documenting the full visual system (color roles, type scale, spacing, radius, motion, z-index, elevation, components, layout).
  • Reworked src/index.css into a token-driven system: every component now consumes design tokens (--fs-*, --sp-*, --r-*, --dur-*/--ease, --z-*, color roles) instead of scattered literal values.
  • Consolidated the typographic scale: the old 8/9/9.5/10px label sprawl collapses to one 10px eyebrow tier, small controls move to a legible 11px, and the wordmark/footer keep Cormorant. Unified eyebrow letter-spacing and tightened the header/wordmark lockup rhythm.
  • Removed a dead parallel design system: deleted six unused components (ActiveMixer, Header, MasterControls, PresetManager, SleepTimer, SoundLibrary) that were the only consumers of the Tailwind theme, then stripped the now-unused Tailwind theme tokens (midnight, deepBlue, cardBlue, accentSoft, hero-gradient, shadow-card, pulseSoft).
  • De-duplicated the slider-fill helper (previously copied in three files) into src/lib/sliderFill.ts.

0.2.1

• Accessibility polish pass (no behavioural changes):
  • Raised muted-text contrast to meet WCAG AA: --text-secondary 3.5:1 → 4.7:1 and --text-dim 1.9:1 → 3.4:1 against the background, keeping the cool blue-grey hue. Section labels and the preset-name placeholder now use the secondary tier so their small text clears 4.5:1.
  • Added a prefers-reduced-motion: reduce block: the moon drift, scroll parallax, star pulse, play-button pulse, equaliser bounce, and entrance reveals all settle to a static state; the equaliser stays visible as a static mark so "playing" still reads.
  • Added :focus-visible keyboard focus rings (accent outline) to every control; previously each set outline: none with no replacement. Card toggles ring the whole card to avoid clipping inside the card's overflow: hidden.
• Removed unused src/components/VolumeSlider.tsx (dead code; the app renders sliders inline).

0.2.0

• Major sound expansion: Added 11 new procedurally generated sounds: Waterfall, Tent Rain, Tin Roof Rain, Underwater, Shower, Frogs, Cafe, Airplane, Dryer, Deep Space, and Heartbeat — bringing total to 25 sounds
• Meaningful sliders: Every sound now has 3 tuning sliders that actually regenerate the audio with different parameters (e.g. Rain intensity changes drop density, heaviness changes filter character, surface changes resonance). Replaced the old playback-rate/gain-multiplier approach with real WAV regeneration on slider change
• New category: Added "Cozy" category with Heartbeat; expanded Water (8 sounds), Urban (4 sounds), Noise (4 sounds), Wildlife (3 sounds)
• Smooth parameter updates: Slider changes trigger debounced WAV regeneration (300ms) with seamless 400ms crossfade to the new audio — no audible gaps
• All existing sounds retain their original character at default slider positions

0.1.18

• Reduce audible loop-gap risk while preserving sound character: keep crossfade timing based on sound end, but add a lightweight playback monitor (120 ms cadence) so crossfades still trigger smoothly even when browser timeupdate events are sparse
• Prevent category chip layout shift: always reserve the count circle space and show an inactive grey placeholder when the active count is zero

0.1.17

• Move version number from header superscript to a tiny fixed label in the bottom-right corner of the screen

0.1.16

• Fix element watermark overlapping volume slider: repositioned to top-right of card with smaller size
• Add "Noise" category (music note icon) for White Noise, Pink Noise, Brown Noise
• Add "Wildlife" category (raven icon) with new Birdsong sound (chirps, trills, and peeps over gentle forest bed)
• Add sound selection buttons to Sound Builder (Fire, Birdsong) for choosing which sound to tune

0.1.15

• Fix persistent bubble/stream character in fire pops: remove all sine-wave components (f1/f2 oscillators at 180–560 Hz were the root cause) and replace with a 100% noise burst; shorten max pop life from 115 ms to 68 ms and tighten envelope (exp(-12p)) for a sharper crack with no tonal tail

0.1.14

• Bake in user-tested optimal fire settings as new defaults (bodyVol 1.4, bodyLp 0.007, roarMean 0.81, roarSpeed 0.00005, roarSigma 0.0015, crackleBase 13.5, crackleVol 5.4, popVol 1.35)
• Fix bubble sound in pop renderer: remove descending pitch chirp and increase noise ratio to 70%; update SoundBuilder default values to match

0.1.13

• Fix build error: add missing src/vite-env.d.ts with /// <reference types="vite/client" /> so TypeScript recognises import.meta.env.BASE_URL

0.1.12

• Fix worklet module path: replace hardcoded /worklets/fire.worklet.js with import.meta.env.BASE_URL + worklets/fire.worklet.js in both useAudioMixer.ts and SoundBuilder.tsx — absolute path was resolving to domain root instead of /SleepSounds/ on GitHub Pages, causing "Unable to load a worklet's module" error and silently falling back to pre-computed audio in the main mixer

0.1.11

• Fix SoundBuilder play button: use a module-level AudioContext singleton and call ctx.resume() synchronously before first await, ensuring it runs within the user gesture on iOS Safari; add visible error display so failures surface in the UI

0.1.10

• Fix SoundBuilder play button: reuse FireWorkletSource shared AudioContext instead of creating a new one, which caused suspend/resume failures on iOS Safari

0.1.9

• Renamed fire sound label back to "Fire"
• Slowed EQ bar animation from 0.9 s to 1.6 s per cycle
• Replaced fire background hiss with a rolling thunder roar: deep LP (~35 Hz) modulated by a slow OU envelope (~0.75 s time constant) for a distant-thunder swell/fade effect
• Crackles louder (mix weight 2.0 → 2.8) and more frequent (base rate 0.5 → 3.0)
• Exposed 8 new real-time AudioWorklet k-rate params on fire-synth processor (bodyVol, bodyLp, roarMean, roarSpeed, roarSigma, crackleBase, crackleVol, popVol)
• Added Sound Builder dev tool: collapsible section at bottom of app with grouped sliders for all fire parameters and a copy-to-clipboard config output
• Reclassified sounds into elemental categories: Water, Fire, Air, Earth
• Added faint element icon watermark to each sound card (Material Symbols); brightens when active

0.1.8

• Rewrote genFire with a 9-layer architecture: deep brown roar, pink body, flame hiss, ember sizzle, air whoosh, clustered crackle bursts with resin pings, spit crackles, pops, and log-shift rumble events
• Fixed fire sounding like white noise — lowered hiss/ember filter cutoffs, boosted roar/body weights, reduced continuous high-freq layer contributions
• Bumped PRESET_STORAGE_KEY to v2

0.1.7

• Initial WAV generation implementation