docs(nook): mark completion type-to-filter shipped, correct the visual-loop note

The 'Depth passes' candidate still listed type-to-filter as open and
repeated 'wants a visual loop, not a blind hour' — both stale since it
landed in a128fa0. Move it to an Unreleased ledger entry, retarget the
next completion slice to matched-character highlighting (which genuinely
wants a visual loop, for the lipgloss nested-background-bleed reason now
documented), and add a harness note recording that host-view slices are
verifiable headless via newModel + Update + View.

Author: truffle-dev

docs/nook/roadmap.md

@@ -72,6 +72,19 @@ Grouped by arc. One line per capability; the tag is where it landed.
   switcher, settings overlay, LSP-ordered completions) and the
   `multi-select` component completing the form-input family.
 
+### Unreleased (on main, not yet tagged)
+
+- Completion type-to-filter (`a128fa0`): typing an identifier rune while
+  the completion popup is open now narrows the visible list live against
+  the trailing word prefix (case-insensitive subsequence, so "wg" keeps
+  "WaitGroup") instead of dismissing the popup. Backspace widens back
+  without a fresh server round-trip; deleting the whole prefix or typing a
+  non-identifier rune dismisses and falls through to the buffer. This is
+  the Zed/VSCode contract. `complete.Popup` retains the full sorted server
+  set (`all`) so `Narrow` can widen without re-querying; the host arms a
+  `completeNarrowPending` flag on the keystroke, lets the rune reach the
+  buffer, then recomputes `WordPrefix` from the live line and narrows.
+
 ## Standing constraints (do not regress)
 
 These are operator-stated DNA. They override convenience.
@@ -91,6 +104,23 @@ These are operator-stated DNA. They override convenience.
    components in v0.48.0 had to backfill entries into the release commit
    because they didn't; toggle did it right.
 
+## Harness note: host-view slices are testable headless
+
+A recurring miscall in this file was tagging keystroke-routing plus
+live-render work as "wants a visual loop, not a blind hour." The
+completion type-to-filter slice disproved the blanket version of that:
+the host model is fully drivable in a unit test. `newModel(dir)` (or
+`fixtureRepo(t)` + `newModel(root)`), set `width`/`height`, `m.resize()`,
+`openBufferForTest(t, m, "a.go")`, then feed `m.Update(tea.KeyMsg{...})`
+and assert on `m.View()` — the full rendered frame — or on sub-pane
+state. The narrow tests check exactly this: type 'i', assert the popup
+survives and the frame drops the filtered-out item. So a live-render
+change is a blind hour only when the thing being rendered is invisible to
+`stripANSI` assertions — the matched-highlight ANSI-bleed case above, or
+anything whose correctness is a color/background composition. Structural
+render changes (which rows appear, what text they carry) are verifiable
+headless; reach for the harness before deferring to a visual loop.
+
 ## Release ritual (reusable)
 
 1. Release commit updating `CHANGELOG.md` (move `[Unreleased]` into the
@@ -132,12 +162,19 @@ shipped, so the remaining surface is narrower than it once was.
 - **Settings/themes surface** — config inheritance and theme hot-reload
   ship; a discoverable settings UI on top does not yet.
 - **Depth passes on shipped panes** — diagnostics presentation and
-  multibuffer ergonomics are each self-contained slices. Completion has a
-  concrete gap: the popup sorts by LSP SortText but has no type-to-filter
-  — typing any rune while it is open dismisses it (main.go, the "any
-  other key" branch) rather than narrowing the list live the way Zed and
-  VSCode do. Closing that is a host keystroke-routing plus live-render
-  change, so it wants a visual loop, not a blind hour.
+  multibuffer ergonomics are each self-contained slices. Completion's
+  type-to-filter gap is now closed (see Unreleased above). The next
+  completion slice is **matched-character highlighting**: bold or accent
+  the runes in each label that the current prefix matched, the way Zed
+  underlines the fuzzy hits. That one genuinely wants a visual loop, and
+  for a specific reason — the row already renders through a lipgloss style
+  that sets a `Background` (`rowStyle` uses `Surface`, `selStyle` uses
+  `Primary`). Embedding a per-rune inner style inside that line emits an
+  ANSI reset mid-line that clears the outer background for the rest of the
+  row, a bleed a `stripANSI`-based headless test cannot see. The correct
+  build composes the highlight into the same style pass (or pre-segments
+  the row so no inner reset lands before the background is re-applied) and
+  eyeballs the rendered frame. Do not ship it blind.
 
 Pick the smallest finishable slice that advances parity; one capability,
 its tests, its CHANGELOG line, one commit.