diff --git a/unified-design-system/nielsen-audit.md b/unified-design-system/nielsen-audit.md
index 717bb51..f6fcf38 100644
--- a/unified-design-system/nielsen-audit.md
+++ b/unified-design-system/nielsen-audit.md
@@ -14,6 +14,11 @@ more repos already contain a structurally similar implementation**.
 
 Written 2026-08-14, alongside `design_system` v0.1.0.
 
+**Phase 3 (web, `@kuhyx/web-ui`) is recorded in `phase3-record.md`** — including
+why `FilterBar` was *not* extracted (structurally disjoint despite the shared
+name) and the new categorical ramp. Read it before re-proposing any web
+component.
+
 ## Scope
 
 Fourteen Flutter repos, seven web repos, four Python/Tk guard apps. Out of
diff --git a/unified-design-system/phase3-record.md b/unified-design-system/phase3-record.md
new file mode 100644
index 0000000..b3a1968
--- /dev/null
+++ b/unified-design-system/phase3-record.md
@@ -0,0 +1,113 @@
+# Phase 3 record — `~/utils/web_ui` (`@kuhyx/web-ui`)
+
+The web counterpart to Phase 1's Dart `design_system`. Companion to
+`nielsen-audit.md`, which holds the grid and the deferred clusters.
+
+Shipped as `web_ui-v0.3.0`, consumed by **dufs-cloud** and
+**awesome-mcp-explorer**.
+
+## What shipped
+
+| Item | Notes |
+| --- | --- |
+| `tokens.css` | Full palette, spacing, radius, type, shadow tiers, focus ring, and the global `:focus-visible` rule. |
+| `RangeSlider` | Reconciled from the two drifted copies; keyboard-operable. |
+| `quantile.ts` | `nth`, `clamp01`, `quantileValue`, `valueQuantile`, `fractionFromPointer`. |
+| `fuzzyMatch` | Was `fuzzyMatch` in dufs-cloud and `matchesFuzzy` in amx. |
+| Categorical ramp | `--cat-1…6`; see `tokens.md`. New token group, see below. |
+
+## FilterBar was NOT extracted — and should not be
+
+The phase brief listed `FilterBar` + `filter-sort.ts` as the third component,
+to be "reconciled" between dufs-cloud (174 lines) and awesome-mcp-explorer
+(225). Inspection says they are a **name collision, not a duplicate** — exactly
+the Phase 1 trap the brief itself warns about, one level deeper than the three
+`FilterSheet`s.
+
+They share a name and nothing else:
+
+| | dufs-cloud | awesome-mcp-explorer |
+| --- | --- | --- |
+| Shape | horizontal `<div class="filterbar">` strip | vertical `<aside class="filters">` sidebar |
+| Controls | search input, `<select>`s, sort key + direction | tri-state chip pickers, no search, no sort |
+| Domain | files (`DirEntry` + `MetaIndex`) | MCP servers (`Server`) |
+| Props | `onFilter` + `onSort` | one `onChange` |
+| `FilterState` | `type`/`extIncludes`/`minSize`/`minDurationMs`/`minPixels` | `languages`/`scope`/`os`/`foss`/`cost`/`grades` |
+| `SortKey` | 9 file keys | disjoint server keys |
+
+`applyFilterSort` differs in arity and parameter types. Extracting a component
+over this would mean inventing a generic that neither caller wants, then
+threading both vocabularies through it — the "shipping a widget nobody imports"
+failure mode, with a forced abstraction on top.
+
+What *was* genuinely duplicated inside those files was extracted: the fuzzy
+subsequence matcher (identical algorithm, renamed identifiers) and the whole
+`quantile.ts` module (near-verbatim, comments aside).
+
+**Do not re-raise FilterBar as an extraction candidate** without new evidence —
+specifically, a third repo whose filter surface is structurally like one of
+these two. A shared *sort-direction toggle* or *filter-chip* is a plausible
+future extraction; the bar itself is not.
+
+`passesTri`/`TriSelect` was checked as a possible third duplicate: dufs-cloud
+carries `extIncludes`/`extExcludes` in its `ExtensionPicker`, but the logic is
+inline and file-extension-specific, not a shared generic. One repo is not two,
+so it stays in amx.
+
+## The categorical ramp (new)
+
+The brief flagged that the design system has one accent and therefore no answer
+for "N mutually distinguishable hues", and said to raise it rather than invent
+one inline. It was raised; the decision was to **design a real ramp** rather
+than map grades onto `success`/`warning`/`danger` (which would have collapsed
+six ordinal steps into three and made A and B identical).
+
+Six colours, chosen by search under measured constraints, documented in
+`tokens.md` and re-checkable with `scripts/ramp_check.py` (exits non-zero on a
+regression). The binding constraint turned out to be theme independence: a
+colour readable on both `ink` and `paper` must sit in L\* 44–59, which is why
+the ramp is mid-toned throughout. A seventh hue is not available at ΔE 20 — the
+best candidate reaches 17.9.
+
+amx's grades now map onto it. **This changed those colours visibly**, and for
+the better: the old hand-picked A/B pair (`#2f8f5b` green vs `#4f8f2f` olive)
+was nearly indistinguishable under deuteranopia.
+
+## Follow-ups
+
+1. **Dart side does not have the ramp.** `tokens.md` is the source of truth for
+   both stacks and now documents `cat-1…6`, but `~/utils/design_system` ships
+   only the original palette. A `design_system` bump should add it, so the
+   "same hex values everywhere" invariant holds. Deliberately not done here:
+   Phase 1 is tagged with six consumers, and that retag is its own change.
+2. **Python/Tk side** likewise has no ramp (no consumer needs one yet).
+3. **npm cannot consume this package.** `utils` has no root `package.json`, and
+   npm's `github:` shorthand has no subdirectory syntax — it clones the root and
+   fails with `ENOENT` (verified, npm 11.16). pnpm's `&path:/web_ui` works, so
+   both consumers are pnpm. A future npm-only consumer needs either its own repo
+   or a published tarball.
+
+## Verification performed
+
+- `pnpm test` + 100% coverage in all three repos (web_ui 84, amx 405, dufs 224).
+- Built bundles grepped: shared tokens present, every bespoke value absent —
+  definitions *and* `var()` references, since checking only definitions lets a
+  shadowed token survive.
+- Real headless Chrome at 1366x768 and 1024x600: `scrollWidth <= innerWidth`,
+  and `:focus-visible` computed as `2px solid rgb(184,134,46)` after a real Tab
+  (a programmatic `.focus()` does not match `:focus-visible` and will report no
+  ring even when one is correctly styled).
+- dufs-cloud checked against a live `dufs` backend, since a static file server
+  answers `501` to `PROPFIND` and the app never renders its filter bar.
+
+## The bug jsdom could not catch
+
+The first cut of the keyboard support stepped by *array index*. Every unit test
+passed, on an evenly spaced ten-value fixture. In the browser the arrow keys did
+nothing: ~1500 of amx's ~2981 star values are `0`, so a step inside that run
+moved the index and left the value — and therefore the thumb, the readout and
+the filter — exactly where it was.
+
+`steppedValue` now walks to the next *distinct* value. The lesson is about
+fixtures, not about arrows: an evenly-spaced fixture is not a distribution, and
+the component's whole reason for existing is that real distributions are lumpy.