What shipped.

Four rounds took the registry from 439 to 534, sourced across twenty-four axes and built by parallel agent teams, with every component judged by hand afterwards. The axes were textile and fibre process, machining and metrology, analog recording mechanism, reprographics, recording instruments that draw on moving media, untaken scan and raster technique, laboratory instrumentation, grain and milling, glassworking and kiln, mechanical cipher and computing, optics manufacture, joining and fastening, horology and escapements, ice and cryo, pneumatics and fluid power, gas discharge, mail and material handling, fermentation and colloids, drilling and rock handling, built animal structures, railway permanent way and signalling, leather and cordage craft, optical telegraphy and beacons, and soil and granular flow. Round 9 shipped all twenty-eight; rounds 10, 11 and 12 lost nine, six and eight to owner review respectively, all quarantined rather than deleted.

Sixteen components took the registry from 423 to 439, sourced on four axes rather than round 8a's two: glyph and character-cell rendering technique (NAPLPS separated mosaic, row-multiplexed LED panel, sprite multiplexing, vector-display beam retrace), print and bindery process (rotogravure doctor-blade wipe, risograph multi-pass registration drift, screen-print flood-and-stroke, web-offset ink-train cascade), industrial and fluid process (peristaltic pump occlusion, shot-peening coverage saturation, gas-fluidized bed, bubble-cap distillation tray), and display-hardware artifacts (e-ink waveform ghosting, LCD overdrive overshoot, CMOS rolling shutter, projector gate weave). All sixteen passed the automated gate on the first review pass, and fourteen passed owner review untouched. The two that came back were both legibility failures rather than mechanism failures: meter-matrix-scan's 240Hz row-scan clock aliased against the 60Hz paint rate and read as a strobe rather than as multiplexing, fixed by decoupling the visual into a 7.5s soft sweep while keeping the PWM boundary-column quantization that is the component's actual identity; and overflow-chip-mux rotated its slots every 130ms, which is faster than anyone can follow, now 1100ms with an explicit leave-then-arrive rather than a blink. The client JS budget was re-baselined from 1778 KB to 2033 KB, and the growth was traced before accepting it: the largest chunk is the registry metadata index carrying every component's description and instruction prose, not component code, so the honest upgrade path is to stop shipping the instruction field to the client rather than to keep raising the ceiling.

Thirty-four components landed, taking the registry from 389 to 423. Two sourcing axes, both grounded in observed mechanics rather than invented categories. The first: glyph rendering TECHNIQUE. The registry's ~62 existing ASCII slugs all shared one pattern — simulate a scalar field, map luminance through a ramp glyph, keep most cells empty — so "which field to simulate" was exhausted and the open axis was the technique itself: sub-cell quadrant blocks, braille 2x4 dot addressing, Sobel gradient ANGLE rather than magnitude, Floyd-Steinberg propagated error as against the existing ordered-dither family, void-and-cluster blue noise, PETSCII binary reverse video, sixel raster-versus-glyph contrast, teletext 2x3 sextants, typewriter overstrike, dot-matrix NLQ half-pitch, mezzotint (the registry's only subtractive mechanic), ZX Spectrum attribute clash, ClearType subpixel slivers, contact-screen dot gain. The second: real physical and industrial process — SMPTE film leader, three theatrical curtain riggings, same-ink halftone moire, rose-engine guilloche, capillary wicking, Jominy end-quench, green-sand shakeout, glaze bubble-and-heal, spark testing, longwall mining, Kelvin wake, crack-arrest stop-drilling, Faraday waves, thin-film telephone-cord buckling, axial slice-comb, text-as-geometry, silhouette glyph packing. Five components were cut on review after passing every gate: a storage tube, a cloud chamber, a gauge-block wringing panel, an eddy-current brake and a CRT glass wrapper — all built, all working, none earning their place. Then the owner looked at a card on a Retina display and found curtain-leader-countdown drawing itself past the bottom-right corner. A canvas is a replaced element, so absolute inset-0 with no width/height CSS falls back to its intrinsic size: at dpr 2 the CSS box became 4000x1800 inside a 2000x900 container, and at dpr 1 it coincidentally matched, which is why every automated check passed it. The audit that finding triggered turned up five more, each invisible to the gates for its own reason. empty-state-braille-orbit clamped its arc-length march to the per-frame time budget, forcing one sample per frame spaced by wall-clock time so the arc-length control never engaged at all, rendering a wandering four-dot fragment instead of an orbit. All three fabric curtains declared autoplay: press, and the embed card's driver fired a real click that opened them permanently — so the thumbnail recorded the revealed empty page rather than the drape, a bug verify.ts structurally cannot see because it navigates without the autoplay parameter. shearer-advance's demo was a one-shot state machine that filled to MAX_ROWS and stopped, freezing its card pixel-identical from t=13s onward. Three backgrounds had content-protection mechanisms that did not work at card size: the guilloche "portrait window" was a circle inscribed in min(w,h), starved by the shorter axis on a wide hero, and the canopy's protective origin bias sat inside its own content block, planting the trunk under the copy it existed to avoid. And two components rendered correctly but framed themselves badly — a grinding wheel pinned to a corner with 60% of its card empty, a ship wake filling its strip edge to edge. Hardening from the same sweep: shakeout-crumble's terminal committed state survives autoplay only because its period exceeds its arm window, which is now documented as load-bearing rather than left as a coincidence. docs/review-workflow.md records the whole review and verify recipe — the three-route split, the launch steps, the gate's hittability test, and every failure mode hit along the way — so the next round does not rediscover them.

Sixty-three components landed across six lanes — identity/access & trust (12), money in motion (12), multiplayer/presence & annotation (12), living systems & growth (12), wayfinding & spatial (9), reliability & operations (6) — taking the registry from 326 to 389. Two more, sear-notch and blowdown-seat, were built, passed the gate, and got cut anyway: both worked, neither could be read. Component pages now carry sidebar prev/next navigation with a position counter, instead of leaving Newest-sort as the only way to move between them. A reduced-motion audit swept 117 suspects and found four components suppressing state-change feedback rather than vestibular motion — gauge-capacity-waterline had no passive readout at all, so its value changed with nothing on screen to show it; the other 107 were already correct and untouched. Separately, chart-bar-halftone, chart-donut-halftone and loader-thread-spool were framing their catalog cards on <SmoothCursor>'s own <svg>, injected into every route ahead of any demo's markup, and rendering blank — three of the four components that shipped this way, and three of those were live in production before it was caught. card.focus now has to be a scoped selector: build-autoplay.ts fails the build on a bare tag name, naming the component and the actual failure (a bare tag matches global layout chrome before the demo's own element, not just "invalid selector"), and the twenty other bare-tag values already in the registry — mostly canvas and button, verified harmless only because svg happens to be the one tag the layout injects unconditionally — were converted to scoped selectors rather than grandfathered in, so the next global element the layout gains doesn't reopen the same hole silently.

The mobile nav drawer closed on Escape and backdrop click without returning focus to its toggle, leaving a keyboard visitor's focus on an element that had just gone -translate-x-full invisible; it now restores focus the same way command-palette.tsx already did. That palette declared aria-modal="true" but only handled Escape, Arrow and Enter, so Tab walked focus straight out into the sidebar links behind the overlay while the dialog stayed open — Tab is now trapped on the one real focusable control, the input. Both search inputs, the palette's field and the sidebar tree filter, were outline-none with no focus-visible replacement, so focusing them produced no visible change at all; globals.css had already been written assuming a ring that nothing supplied. Separately, the testimonial submission cap counted only rows still status: "pending", so an owner rejection cleared the limiter and the same account could resubmit immediately — volume was bounded by review speed, not the 24h window. It's now durable and status-independent in a new testimonialRateLimits table. validateSubmission, the Convex-side enforcement point a direct caller reaches by skipping the HTTP route, never checked company, and neither role nor company had a length bound anywhere; both now share MAX_NAME_LENGTH. The status-snapshot cron compared CRON_SECRET with a plain !==; it now uses the constant-time helper already in the codebase. And the rate-limit message that promised "a submission awaiting review" — untrue once the cap stopped counting pending rows — now states the actual rule. Riding along: the changelog page's release tide renders instead of an empty axis, and all 1524 prop rows across the registry's 295 components now carry a description, up from 706.

The ⌘K trigger had been getting scaled up for two rounds running — bigger font, bigger padding, same small chip — and still read as something that got inflated rather than designed. It's now a proper full-width search field in the sidebar header: magnifier, muted "Search components…" placeholder, the ⌘K hint right-aligned in its own kbd chip, opening the same palette it always did. The sidebar's own tree filter, which used to sit in an identical bordered box one row below and had already caused visitors to assume the two searched the same thing, is now visually subordinate — borderless, a funnel glyph, grouped with Expand/Collapse instead of dressed as a second search bar. Every one of the 298 component pages was missing the theme toggle entirely before this pass; it's in the shell now, alongside a skip link and prev/next navigation between components. The footer is regrouped into four columns instead of one long run. Smooth scroll (Lenis) and a custom cursor now run site-wide rather than on isolated pages. /preview/<slug>/play is gone, folded into /components/<slug> with a permanent redirect from the old address, and every component demo now renders inside one shared framed stage — the fixed 520px well that used to clip taller heroes is gone with it. Canonical tags, noindex on the pages that shouldn't rank, seven dedicated OpenGraph images and an RSS feed for /writing are new. Click targets were enlarged across 18 routes, and this is the first pass at a mobile layout below 1440px. /install and /theming are new documentation routes, and the MCP server's token list is now generated straight from globals.css instead of hand-maintained. Saved components show live previews instead of static placeholders, and props tables show real default values — 196 of 298 components had been silently missing them. An Ask AI launcher now sits on /connect, every component page, the footer and the homepage; its own popup auto-dismisses and stops appearing at all after two views instead of nagging indefinitely. /status can no longer fail a production build over a slow CDN response. Copy buttons were added to descriptions, build specs, props, feed URLs and changelog entries themselves.

Every component in this registry was styled against tokens the registry never shipped, so npx shadcn add produced components that rendered wrong in someone else's project and gave no error doing it. --muted and --accent also collided semantically with stock shadcn — theirs is a light background and a subtle grey surface, ours was body text and electric blue — so installed body copy came out nearly invisible. Both are renamed to --ns-muted and --ns-accent, and 266 of the 298 registry items now ship a cssVars block so the tokens arrive with the component; --background, --foreground and --border are deliberately left to inherit from the host theme, because a component that brings its own background looks foreign in every project it lands in. The registry grew from 266 to 298: eight ASCII fields built on real mechanisms — Lichtenberg discharge, Schlieren density gradients, shock diamonds, reaction-diffusion fronts, Voronoi and magnetic domain walls, force chains, nodal lines — plus ASCII instruments that do a job (a git graph, a log tail, a flamegraph, a kanban board with WIP limits, a Gantt with a real critical path). A new /status page reports what is actually true rather than what is reassuring: it shows no uptime figure until it has recorded days to compute one from, it says "not measured" where it cannot honestly check, and it refuses to claim sign-in works because a public query resolving proves only that the backend can be read. Every failure this registry has ever had returned HTTP 200, which is why a conventional green banner would have been a lie. Behind it, monitoring now samples every ten minutes and a day's bar is derived from all of that day's samples, so one lucky ping cannot paint a bad day green. Convex now deploys as part of the Vercel build — the drift that once left /submit and every account feature returning 200 while completely dead has happened three times, and it is now structurally impossible rather than merely documented. All 72 preview videos were regenerated: every one had been named for a slug that no longer existed, so no featured card had ever moved. The screenshot gate went from covering 266 of 298 components to all 298. Five automated gates now guard the things that used to depend on remembering — focus rings, disabled states that still react to the mouse, animations that ignore reduced motion, unreachable categories, stale CSS variables, and functions that exist in the repo but not on the deployment — and each one was deliberately broken first to prove it fails.

Every install command published in the CLI and MCP packages had been returning 404. The 223-slug rename shipped after those packages were built, not before, so both were handing out addresses that no longer existed — and because a failed npx shadcn add happens in someone else's terminal, nothing here would ever have reported it. Every old address now redirects to its current one, and both packages have been republished with correct data, so old installs keep working and new ones resolve directly. The registry grew from 228 to 266 components, most of them ASCII instruments: dithered bar, line, funnel, radar, scatter, waterfall and box-plot charts, a choropleth, sankey, treemap and flow diagram, a patchbay, seatmap, keymap and spreadsheet range, plus new ASCII backgrounds and heroes. The catalog's own "Newest" sort had been putting new components last instead of first, which is fixed. Scrolling the catalog could flip the whole site between light and dark: the theme-toggle demo shared a storage key with the site itself, and autoplay pressed it as you scrolled. The homepage no longer shifts its layout while it loads. /community, /guidelines and /submit are new: the first two are written, and /submit opens a pull request against this repo under your own GitHub account, which needs one more callback URL registered before it can complete. The screenshot suite that guards every component was skipping hover, press and focus entirely on 18 of them — anything whose main control is an input, slider, checkbox or radio — and closing that gap immediately turned up a text field with no visible keyboard focus state.

You can now sign in with GitHub, Google or an emailed code and save components to your account. Signing in stays entirely optional: npx shadcn add, the CLI and the MCP server all work with no account, and the catalog is unchanged for anyone browsing without one. Component pages moved from /preview/<slug> to /components/<slug>, while the install URLs at /r/<slug>.json are untouched, so every install command anyone has already run still works. Those pages used to be the live demo and nothing else, and now carry the component's name, description, its own install command, a props table, dependencies and tags, which is also why all 228 components are in the sitemap for the first time with structured data behind them. Next.js moved from 16.2.10 to 16.2.11 for a security fix. Deleting an account is specified but not built yet.

Tabbing the grid drew a focus ring in the corner of a card while a mouse got the whole card as one hit target, the same accessibility mismatch on the site's primary content. The ring now lives on the element hover already highlights, verified with real Tab key events rather than a bare .focus() call, which would not have caught it. Five smaller polish defects went with it: the footer stopped spreading three unrelated lines edge-to-edge at desktop width, mobile filter chips became a two-column grid instead of one per row with dead space beside it, the sidebar's component search and the catalog's own search now say which is which since they do different jobs, the Sort select wears the site's own chevron instead of the browser's, and card hover gained a perceptible wash since brightening a hairline border was invisible against a near-black demo. docs/polish-audit.md records every item considered, including one left alone on purpose.

Category now opens to kind (Slider, Loader, Tabs) and kind opens to components, each level counted, because a flat 223-item list under 12 categories was still a list. A kind only earns its own group at two or more members, so a category with 35 kinds and 25 singletons doesn't add a click to reach a third of the registry for nothing. The count next to each category in the sidebar had drifted from the number on the matching filter chip, 61 versus 55 for Inputs & forms, because the tree filed a component under its first matching category while the chips counted every match. Both now use the same multi-match rule, verified sidebar against chip across all twelve categories.

Component names kept their metaphor at this point, renaming 223 of them was a separate decision for later, but each now carries a plain "kind" caption derived from the tag each meta.json already carries: Nested Slug reads as Nested Slug / Hero, not something you click through to understand. The playground gained a Source disclosure showing the real component.tsx with a copy button, since these install as plain files with no runtime package, so the source is the artifact. And a stdio MCP server shipped alongside the CLI: five tools, search, full detail plus source, categories, the install command, and the design-token conventions, so an agent building against the registry keeps the token contract in context for a session instead of re-fetching it.

tide-ledger, a calendar heatmap on a five-step sequential ramp mixed at runtime from --accent into --background, joined the registry. Two siblings built alongside it in the same pass, gel-wash (a theatre-lantern preloader) and footing-course (a footer the page slides off), came out again: both were built from a bookmarked slug rather than from the thing the slug was pointing at, so they answered a category rather than the actual gap. The backlog behind all three is now one organized document across the 21st.dev bookmark sweep and a name-level diff against seven other component ecosystems, ordered by what unlocks a whole page rather than what is easiest to build next.

The repo got the scaffolding a public project needs: CI running the registry build, typecheck and production build on every pull request, issue templates for bugs and component requests, a pull request template, a code of conduct, a security policy, and a contributing guide that walks an outside contributor from an empty folder to a passing verify gate. The README was rewritten around what the registry is for rather than how it was made, and the agent-only scaffolding used to build it started coming out of the tree.

Two defect classes only show up where nobody looks. One is a component that reads correctly in dark and collapses in light. The other is a grid thumbnail framed on the wrong element, cropping away the very thing the component exists to show. All 197 were swept for both. Ten card-crop cases were fixed by giving each a card.focus selector aimed at the element the component is actually about, checked on the homepage grid rather than on the full preview route where the defect is invisible.

Twenty-eight components merged from a fourth parallel round, 21 core and 7 loud, among them a Solari departures board, a peelable decal, a plimsoll gauge and a vortex street. That took the registry to 198, and tack-peel came out for good, leaving 197. Several older components got the second look they needed: umbra-toggle's controlled state had drifted out of sync and its demo was inert, decal-peel inverted its gradient in the light theme, carbon-lift's ghost duplicate travelled a fixed 12px regardless of how far it actually had to go, solari-flap grew from one row to a four-row board, and patina-pip and stipple-year were made legible at card size.

The parallel round produced more than it deserved to keep. Fifteen slop and duplicate components were removed, taking the registry from 185 to 170, and one that a merge resurrected had to be cut twice. What remained went through four rounds of owner review: ten flagged components repaired, then 17 interaction bugs, then 8 more including reworks of earlier fixes, then a ridge-walk redesign and six more reworks. Three hardcoded hexes were found and killed. The verifier stopped aborting the whole sweep when a single component threw, and started printing every failure rather than only the first.

Six git worktrees, each its own branch and dev port, ran the same ideate-judge-build-gate workflow across a different lane of the design space, then merged back conflict-free because every component is a self-contained folder and the registration index is generated, not hand-edited. The registry crossed 185 components spanning inputs, data instruments, typography, feedback, motion, agent surfaces and loud showpieces, with slug collisions prevented by a cross-branch check rather than coordination.

A first set of AI and agent-interface primitives landed: a thinking-state glyph that encodes its state through motion rather than colour, a streaming-text renderer, a reasoning timeline, a tool-call approval row, a context-window budget meter and more. The autoplay driver learned a type mode, so keyboard-first components like the OTP reels now type their own demo in the grid instead of resting on a still frame.

The grid got a client-side search over name, title, description and tags, composed with the collection filter. The header install command now shows a real component name instead of a [name] placeholder, so what lands on the clipboard actually runs. This page is itself the strandline component from the registry, fed real releases.

A shared autoplay driver synthesises pointer, scroll, press and drag input from an autoplay descriptor in each component's meta.json. 37 components that previously sat still until touched now demonstrate themselves in the grid.

The registry origin became a single source of truth, so llms.txt and the showcase can no longer drift apart. /registry.json moved to its conventional path and llms.txt learned to disambiguate. A real light/dark toggle landed with a system-preference default and no first-paint flash.

Every card is the real component running, not a screenshot. Emulating a viewport with a scaled div drifted at every window shape, so each card is now an iframe onto a real 1440x900 viewport, CSS-scaled into place.

MIT licensed, published, and given llms.txt plus llms-full.txt so an agent can consume the whole catalogue without an MCP server. TypeScript was pinned back to 5.x after v7 shipped no lib/typescript.js and broke the Next build.

24 confirmed bugs and 8 broken light themes fixed across 17 components, then two rounds of owner tuning across thirteen more. Nothing new shipped; everything already there got honest.

A 56-site research harvest and a 14-site crawl catalogued 1,535 components and fed a breeder workflow. Four fusion batches took the registry from a handful to 50, with an ambient-motion pass across all of them.

Spec, owner taste profile, ticket breakdown, then the first working loop: a preview site, a registry build pipeline generated from meta.json sidecars, and a verify quality gate. glass-button and the particle-hero flagship were the first two components in.