tuile 0.12.0 → 0.14.0

data/DECISIONS.md

@@ -13,10 +13,15 @@ It is the *why-we-chose* record; it is not the *how-it-works* reference
 those, put it there and don't restate it — an entry here links out rather
 than duplicating.
 
-**Format.** One entry per decision. The ID is a slug, not a number: `D-`
-(says "this is a decision") plus a 1–4-word kebab hint at the subject
-(`D-bg-inherit`), so a reference carries meaning on its own — a running
-counter would not. The `(date)` on the heading is *decided* provenance,
+**Format.** One entry per decision. The ID is a slug, not a number: `D_`
+(says "this is a decision") plus a 1–4-word hint at the subject
+(`D_bg_inherit`), so a reference carries meaning on its own — a running
+counter would not. **Underscores throughout, never hyphens**: the id has
+to be one *token*, so that vim's `w` / `*` / `ciw` and `grep -w` act on
+the whole thing rather than on a fragment. Backtick it in prose, both
+because some downstream Markdown parsers italicise intraword `_` and
+because a backticked id is copy-pasteable into a search. The `(date)` on
+the heading is *decided* provenance,
 not a log position; git owns the edit history (consistent with the "No
 history" rule — don't narrate how an entry used to read). Keep each entry
 tight: context, the decision, the alternatives rejected and why, and the
@@ -34,7 +39,7 @@ choice, now sharper or broader). Two things this does *not* license:
   It's the most valuable thing in the file.
 - **A reversed *shipped* decision forks a tombstone, it is not overwritten.**
   When a design was tried, shipped, and then thrown away, leave the old
-  entry as the scar, set its `Status:` to **Superseded by D-<slug>**, and
+  entry as the scar, set its `Status:` to **Superseded by `D_<slug>`**, and
   write the replacement fresh. (The shape of such a reversal: the deleted
   bottom-up `content_size` sizing channel, replaced by top-down layout —
   see AGENTS.md "Layout is top-down".) The line: *refined or extended* →
@@ -42,7 +47,7 @@ choice, now sharper or broader). Two things this does *not* license:
 
 ---
 
-## D-bg-inherit — Background color: fill-the-gaps inheritance (2026-07-23)
+## D_bg_inherit — Background color: fill-the-gaps inheritance (2026-07-23)
 
 **Status:** Accepted; implemented 2026-07-23. Tracks
 [issue #1](https://github.com/mvysny/tuile/issues/1).
@@ -115,21 +120,21 @@ are opt-in"). {Component::Label} already carried its own `#bg` (override-all
 via `with_bg`); it composes with `bg_color` (explicit span bgs survive
 `under_bg`, so `#bg` wins locally), but the two-knob overlap is a wart
 flagged for a later consolidation decision. The theme-token variant that
-surfaced during design landed separately — see `D-theme-ref`.
+surfaced during design landed separately — see `D_theme_ref`.
 
 ---
 
-## D-theme-ref — Live theme references for `bg_color` (2026-07-23)
+## D_theme_ref — Live theme references for `bg_color` (2026-07-23)
 
 **Status:** Accepted; implemented 2026-07-23. Tracks
 [issue #1](https://github.com/mvysny/tuile/issues/1). Relaxes the
-`bg_color`-takes-`Color`-only stance of `D-bg-inherit`, which rejected a
+`bg_color`-takes-`Color`-only stance of `D_bg_inherit`, which rejected a
 built-in `panel_bg` token and deferred the general "themeable color
 property" question.
 
 **Context.** Tracking a themed background meant setting the color *twice* —
 once as a concrete `Color`, and again in an `on_theme_changed` block so it
-survives light/dark flips — for every tinted panel. `D-bg-inherit` deferred
+survives light/dark flips — for every tinted panel. `D_bg_inherit` deferred
 the fix; this is it.
 
 **Decision.** `Component#bg_color` accepts a `Theme::Ref` (built by
@@ -163,7 +168,7 @@ existing chrome accent and tracking flips — concretely
 {Component::ComboBox}'s borderless dropdown, which tints with
 `input_bg_color` (tying it to the field's own well) and would otherwise need
 the very `on_theme_changed`/resolve-on-open boilerplate `Theme::Ref` exists
-to kill. The invariant `D-bg-inherit` actually protects is *the Theme
+to kill. The invariant `D_bg_inherit` actually protects is *the Theme
 carries no global bg/fg field* — and every chrome token is an **accent**
 (`active_bg`, `active_border`, `input_bg`, `hint`), never a global
 background. A `Ref` to one adds no new token and creates no global
@@ -200,7 +205,7 @@ reaching chrome tokens leaves the no-global-bg/fg guard untouched.
 **Consequences.**
 - A `Ref` adds no new token (chrome tokens are all accents; `custom` is
   app-supplied), so it **cannot** reintroduce the global bg/fg token that
-  `D-bg-inherit` and the AGENTS.md theme stance refuse — the two stay
+  `D_bg_inherit` and the AGENTS.md theme stance refuse — the two stay
   orthogonal.
 - Collision precedence is chrome-wins; a `custom` token named after a chrome
   token is shadowed when referenced by `Ref` (harmless, documented on
@@ -214,7 +219,7 @@ reaching chrome tokens leaves the no-global-bg/fg guard untouched.
 
 ---
 
-## D-has-value — Typed value seam (`HasValue`) over String-only (2026-07-23)
+## D_has_value — Typed value seam (`HasValue`) over String-only (2026-07-23)
 
 **Status:** Accepted; implemented 2026-07-23 (`Component::HasValue`, included
 by `AbstractStringField`; first typed consumer is `ComboBox`). Tracks the "do input
@@ -271,10 +276,10 @@ values *and* a uniform seam for free.
 
 ---
 
-## D-combobox — `ComboBox`: composed, typed, filterable-first (2026-07-23)
+## D_combobox — `ComboBox`: composed, typed, filterable-first (2026-07-23)
 
 **Status:** Accepted; implemented 2026-07-23 (`Component::ComboBox`, demoed in
-the sampler). Builds on `D-has-value`, `D-bg-inherit`, `D-theme-ref`.
+the sampler). Builds on `D_has_value`, `D_bg_inherit`, `D_theme_ref`.
 
 **Context.** A text field with a filtering dropdown. The ad-hoc version already
 existed in the sampler's slash-command demo (a `TextField` + a non-modal
@@ -302,18 +307,18 @@ component.
   make value-before-items the normal case rather than a corner. The cost — the
   app owns keeping them in sync, reconciling with a one-line intersection when
   it wants to — is smaller than any framework reconcile step (see the rejected
-  three in `D-checkbox-group`, where the set-valued case forced the question).
+  three in `D_checkbox_group`, where the set-valued case forced the question).
   One rule, two instances: singular here, a `Set` of items in `CheckboxGroup`.
 - **Two values, never conflated.** `value` = the committed selection (changes
   only on Enter/click; sole trigger of `on_value_change`); the field's `text`
   = a transient **query** that filters the list and reverts to the value's
   label on ESC/blur.
 - **Filterable first;** the non-filterable `Select` is deferred (it wants the
-  read-only field behavior `D-has-value` parked for the forms layer).
+  read-only field behavior `D_has_value` parked for the forms layer).
 - **Borderless tinted dropdown** (no `Window`): a bare `Popup(List)` told apart
   from the content by a background tint, `bg_color = Theme.ref(:input_bg_color)`
-  — live-tracked, no `on_theme_changed` hook (leans on `D-bg-inherit` +
-  `D-theme-ref`). A `▾` affordance marks the field; the dropdown flips above
+  — live-tracked, no `on_theme_changed` hook (leans on `D_bg_inherit` +
+  `D_theme_ref`). A `▾` affordance marks the field; the dropdown flips above
   when it would overrun the screen bottom.
 
 **Alternatives rejected.**
@@ -321,7 +326,7 @@ component.
   above.
 - *String value (the display text):* fails identity-across-duplicate-labels,
   the whole reason to prefer a component over `List` + a lookup hash
-  (`D-has-value`).
+  (`D_has_value`).
 - *Store the selected **index** rather than the object* (and clear the selection
   when `value=` gets something not in `items`): the plausible misreading of the
   identity rule, and it breaks the chrome/value split above — replacing `items`
@@ -330,7 +335,7 @@ component.
   *resolution* mechanism, valid only at the instant of a click.
 - *`Window`-framed dropdown:* the border is redundant chrome once a tint
   separates the panel, and costs 2 rows + 2 cols; the tint is what
-  `D-bg-inherit` was built to make solid.
+  `D_bg_inherit` was built to make solid.
 - *`allow_custom_value`* (Vaadin's "typed text not in the list" escape hatch):
   deferred — a custom value is a `String`, reintroducing the String/`T` tension
   at the value boundary; no use case needs it yet.
@@ -347,10 +352,10 @@ component.
 
 ---
 
-## D-integer-field — `IntegerField`: the second typed input, and the composed-field taxonomy (2026-07-23)
+## D_integer_field — `IntegerField`: the second typed input, and the composed-field taxonomy (2026-07-23)
 
 **Status:** Accepted; implemented 2026-07-23 (`Component::IntegerField`). Builds
-on `D-has-value`, `D-combobox`. Its real job was to *validate the `HasValue`
+on `D_has_value`, `D_combobox`. Its real job was to *validate the `HasValue`
 seam* for the case where `value`'s type diverges from the editing buffer:
 `ComboBox` proved the fully-detached case (value ⟂ query), `IntegerField`
 probes the *derived* case (value = a parse of the buffer). Extended 2026-08-02
@@ -367,7 +372,7 @@ moment to settle the input taxonomy while still pre-1.0.
   not reuse: subclassing drags `TextField`'s `String`-typed `text`/`value` seam
   onto the field's public face, next to the real `Integer` `value` as a
   conflicting second seam, and Ruby can't cleanly hide inherited public
-  methods. (Same shape as `D-combobox`; makes `IntegerField` a *simpler
+  methods. (Same shape as `D_combobox`; makes `IntegerField` a *simpler
   ComboBox* — the identical structure minus the dropdown.)
   **The taxonomy is two-sided: compose when the value's type diverges from the
   buffer, subclass when it doesn't.** `Component::PasswordField < TextField`
@@ -388,7 +393,7 @@ moment to settle the input taxonomy while still pre-1.0.
   `cycle_focus` collects stops via `on_tree`).
 - **The converter stays private and hardcoded** (`Integer(t, 10)` / `to_s`),
   exactly as `TextField` hardcodes identity-String. No public `converter=`
-  strategy — that is the future Binder's job (`D-has-value` keeps converters
+  strategy — that is the future Binder's job (`D_has_value` keeps converters
   *above* the field).
 - **Value is a derived parse, fired eagerly.** `value` is recomputed from the
   buffer on read; `on_value_change` fires per keystroke but only on a real
@@ -443,7 +448,7 @@ what Java needs a class for, and `is_a?(HasValue)` is the Binder's marker.
 
 ---
 
-## D-ambiguous-width — Bet on ambiguous-as-narrow; keep the inventory small (2026-07-30)
+## D_ambiguous_width — Bet on ambiguous-as-narrow; keep the inventory small (2026-07-30)
 
 **Status:** Accepted 2026-07-30; describes what Tuile already does, plus one
 new *forward-looking* rule (the inventory discipline) that governs new glyph
@@ -546,7 +551,7 @@ Detect once and swap glyphs, rather than re-deriving widths everywhere:
 
 ---
 
-## D-key-dispatch — Delete `key_shortcut`; scope-wide keys ride the bubble (2026-07-30)
+## D_key_dispatch — Delete `key_shortcut`; scope-wide keys ride the bubble (2026-07-30)
 
 **Status:** Accepted 2026-07-30; implemented the same day. Supersedes the
 shipped capture phase of `ScreenPane#handle_key` — see *the scar* at the end.
@@ -758,13 +763,13 @@ on a per-binding flag).
 
 ---
 
-## D-boolean-fields — `Checkbox`: two-state value, painted extent, ASCII glyphs (2026-07-30)
+## D_boolean_fields — `Checkbox`: two-state value, painted extent, ASCII glyphs (2026-07-30)
 
 **Status:** Accepted; `Component::Checkbox` implemented 2026-07-30. Builds on
-`D-has-value`. The glyph and caption rulings are shared with
-`Component::CheckboxGroup` (`D-checkbox-group`, which scopes the key and hit-test
+`D_has_value`. The glyph and caption rulings are shared with
+`Component::CheckboxGroup` (`D_checkbox_group`, which scopes the key and hit-test
 rulings below to a *standalone* widget) and with `RadioGroup`
-(`D-radio-group`). Tri-state is settled here but **not built**, and this
+(`D_radio_group`). Tri-state is settled here but **not built**, and this
 entry is its only home — see the last section.
 
 **Context.** The first boolean input: one row, `[x] Enable syslog forwarding`,
@@ -791,7 +796,7 @@ will follow.
   don't need is the irreversible direction. What tipped it is *consistency with
   the group components*: a checkable row inside a `List` toggles on Enter, since
   Enter is `List`'s own choose-the-item-under-the-cursor gesture
-  (`D-checkbox-group`, `D-radio-group`). So `[ ] Verbose` flipped on Enter when
+  (`D_checkbox_group`, `D_radio_group`). So `[ ] Verbose` flipped on Enter when
   it sat in a `CheckboxGroup` and did nothing when it sat alone in a form — a
   distinction the user cannot see, and one that reads as a bug in the standalone
   widget rather than as restraint. One gesture set, both shapes, is worth more
@@ -827,7 +832,7 @@ will follow.
   `ellipsize(rect.width)` already equals `ellipsize(extent.width)` in both
   directions.
   **The rule is scoped to a *standalone* one-row field.** A checkable row
-  *inside a list* hit-tests its full width instead (`D-checkbox-group`), and the
+  *inside a list* hit-tests its full width instead (`D_checkbox_group`), and the
   difference is perceptual rather than a relaxation of rigor: with a cursor
   visible and ten rows stacked, the unit the user aims at is a **row**, and a
   row's affordance is its whole width — which is what `List`'s row-wide cursor
@@ -845,7 +850,7 @@ will follow.
   so the two states degrade *asymmetrically* to tofu — checked renders,
   unchecked doesn't, which reads as a bug rather than a fallback) and on **ink
   overflow** (the fallback glyph is drawn wider than its cell in Alacritty —
-  cosmetic, coordinates stay correct; see `D-ambiguous-width` for why that's a
+  cosmetic, coordinates stay correct; see `D_ambiguous_width` for why that's a
   different problem). Locally, three columns is also a bigger click target that
   survives a monochrome terminal, and keeps `region_text` assertions ASCII.
 
@@ -860,7 +865,7 @@ will follow.
   the promise in `CheckboxGroup` would have forced it onto the
   `ListDropdown::Menu` shape — a non-focusable `List` subclass plus
   hand-forwarded movement keys — to protect a guarantee nothing relied on
-  (`D-checkbox-group`). Enter-reaches-your-form is a per-assembly property the
+  (`D_checkbox_group`). Enter-reaches-your-form is a per-assembly property the
   app verifies for its own focusable widgets, not a framework invariant. Still
   rejected, and now moot in both directions: the standalone widget claims Enter
   too, which is what made the two shapes agree.
@@ -892,7 +897,7 @@ will follow.
   (`Screen#refresh_status_bar` can't even reach a leaf field — it consults the
   active `Window` or the top popup's *direct* content.)
 - *A read-only flag:* parked with the rest of the forms-layer axes by
-  `D-has-value`.
+  `D_has_value`.
 
 **Tri-state (indeterminate) — settled, not built.** When it lands it adopts
 **Vaadin's orthogonal flag**: `indeterminate`/`indeterminate=` as a plain
@@ -924,23 +929,23 @@ marker is live-resolved chrome like every other built-in accent.
 
 Deferred because the use case (a partially-checked tree parent) has no home in
 Tuile today. Its first plausible consumer would be a `CheckboxGroup` header row
-— which `D-checkbox-group` declined to build, leaving this unbuilt too; that
+— which `D_checkbox_group` declined to build, leaving this unbuilt too; that
 entry names the forcing function to watch for.
 
 ---
 
-## D-checkbox-group — `CheckboxGroup`: a field composing a `List`, `Set`-valued (2026-07-30)
+## D_checkbox_group — `CheckboxGroup`: a field composing a `List`, `Set`-valued (2026-07-30)
 
 **Status:** Accepted; `Component::CheckboxGroup` implemented 2026-07-30, demoed
-in the sampler. Builds on `D-has-value`, `D-combobox` (the chrome/value split it
-generalizes), `D-integer-field` (the composed-field taxonomy it extends) and
-`D-boolean-fields` (the glyphs, and the two rulings it scopes).
+in the sampler. Builds on `D_has_value`, `D_combobox` (the chrome/value split it
+generalizes), `D_integer_field` (the composed-field taxonomy it extends) and
+`D_boolean_fields` (the glyphs, and the two rulings it scopes).
 
 **Context.** Multi-select from a handful of typed items, one `[x] label` row
 each. The cursor and the selection are genuinely two pieces of state here —
 which is exactly the shape `List` already implements, so the question was how
 much of `List` to reuse and what the value should be. (A single-select group
-*could* have conflated them, and `D-radio-group` records why it doesn't.)
+*could* have conflated them, and `D_radio_group` records why it doesn't.)
 
 **Decision.**
 - **Compose a plain `List`, unmodified.** `CheckboxGroup` holds one as its single
@@ -949,7 +954,7 @@ much of `List` to reuse and what the value should be. (A single-select group
   `lines=` on any change to items/labels/selection, claim **Space** in
   `handle_key`, and toggle from `on_item_chosen`. That one callback covers Enter
   *and* click (`list.rb:209` and `:264`), so there is no `handle_mouse` override
-  at all. This **extends `D-integer-field`'s taxonomy** from "a typed field
+  at all. This **extends `D_integer_field`'s taxonomy** from "a typed field
   composes a `TextField`" to "a typed field composes whatever widget already has
   the interaction" — the tab stop lives on the inner widget, the wrapper is not
   one, exactly as for `ComboBox`.
@@ -969,10 +974,10 @@ much of `List` to reuse and what the value should be. (A single-select group
   end — i.e. the observable order is the user's *toggle history*. Documented as
   unordered so nobody builds on that; `items & value.to_a` is the idiom for
   items order, and the sampler pane uses it visibly.
-- **Items are chrome (`D-combobox`), so `items=` never touches `value`** and never
+- **Items are chrome (`D_combobox`), so `items=` never touches `value`** and never
   fires `on_value_change`. A selected item absent from `items` renders no checked
   row and survives intact.
-- **Two `D-boolean-fields` rulings are scoped, not broken.** A click anywhere on
+- **Two `D_boolean_fields` rulings are scoped, not broken.** A click anywhere on
   a row toggles it (a row's affordance is its full width, which its cursor
   highlight already advertises) while a *standalone* checkbox still ignores its
   blank tail; and Enter toggles here because that is `List`'s choose gesture. The
@@ -980,7 +985,7 @@ much of `List` to reuse and what the value should be. (A single-select group
   `on_item_chosen` only for `line < @lines.size`, so a click below the last row
   toggles nothing.
 - **No header row, no tri-state, no select-all.** A header is the only plausible
-  consumer of `D-boolean-fields`' settled-but-unbuilt `indeterminate` flag, and
+  consumer of `D_boolean_fields`' settled-but-unbuilt `indeterminate` flag, and
   it is also where every policy question lives: which children it governs,
   whether checking it selects all, one change event or N, whether it scrolls with
   the rows. That entry already rules a header *app policy*, so building one here
@@ -998,11 +1003,11 @@ much of `List` to reuse and what the value should be. (A single-select group
   whatever now occupies that index; *re-map by `==`* is the honest one but still
   can't preserve intent across duplicates and must decide whether to fire; *clear*
   discards the user's work when items merely gained a row. Storing items deletes
-  the question rather than answering it — see `D-combobox`'s matching rejection.
+  the question rather than answering it — see `D_combobox`'s matching rejection.
 - *The `ListDropdown::Menu` shape — a non-focusable `List` subclass, focus on the
   wrapper, movement keys hand-forwarded:* the design forced by taking Enter away
   from the list. Correct, and about 15 lines of forwarding plus a subclass, all
-  to protect a promise nothing relied on (see `D-boolean-fields`' rejected Enter
+  to protect a promise nothing relied on (see `D_boolean_fields`' rejected Enter
   reservation). Reach for it only if a driver genuinely needs Enter for itself.
 - *Paint the rows directly (`< Component`, `draw_text` per row):* wrong here.
   The cursor-distinct-from-selection structure *is* `List`, a checkbox group is
@@ -1010,7 +1015,7 @@ much of `List` to reuse and what the value should be. (A single-select group
   re-implementing the cursor, the viewport, the scrollbar and the mouse
   arithmetic. This was left explicitly open for a radio group, on the grounds
   that three rows and a selection-follows-cursor model would need almost none of
-  it; `D-radio-group` then closed it the same way, because dropping that model
+  it; `D_radio_group` then closed it the same way, because dropping that model
   removed the friction that made painting attractive.
 - *An `Array`-valued `value` in `items` order:* would make ordering meaningful and
   so make it a contract to maintain, plus `==` would then treat two identical
@@ -1022,7 +1027,7 @@ much of `List` to reuse and what the value should be. (A single-select group
   because that rule is `ComboBox`'s already (the `cop` duplicate-rather-than-fold
   rule).
 - *Public `CHECKED`/`UNCHECKED` glyph constants shared with `Checkbox`:* declined
-  again here for the reason `D-boolean-fields` gives — the group paints its own
+  again here for the reason `D_boolean_fields` gives — the group paints its own
   rows and never instantiates a `Checkbox`, so importing a constant would read as
   a dependency that isn't there. Drift between the two copies surfaces as a
   `region_text` mismatch, not a silent bug.
@@ -1040,12 +1045,12 @@ independent.
 
 ---
 
-## D-radio-group — `RadioGroup`: cursor roams, selection commits; the cursor is chrome (2026-07-31)
+## D_radio_group — `RadioGroup`: cursor roams, selection commits; the cursor is chrome (2026-07-31)
 
 **Status:** Accepted; `Component::RadioGroup` implemented 2026-07-31, demoed in
-the sampler. Builds on `D-has-value`, `D-combobox` (the chrome/value split),
-`D-integer-field` (the composed-field taxonomy), `D-checkbox-group` (the
-`List`-composing shape it copies) and `D-ambiguous-width` (the glyphs). Most of
+the sampler. Builds on `D_has_value`, `D_combobox` (the chrome/value split),
+`D_integer_field` (the composed-field taxonomy), `D_checkbox_group` (the
+`List`-composing shape it copies) and `D_ambiguous_width` (the glyphs). Most of
 this component was settled by those five; what it owns is the **interaction
 model**, which reverses both the desktop convention and this note's own first
 design.
@@ -1114,12 +1119,12 @@ list floors at 0 and a `Cursor::Limited` keeps its own notion of "last". The
   note held open:* it existed to escape the four frictions above, which the
   interaction model removes. Composing a `List` then costs nothing and keeps the
   cursor, viewport, scrollbar and mouse arithmetic in one place.
-- *A `glyphs=` knob for `(•)`:* `D-ambiguous-width` blesses an opt-in knob but
+- *A `glyphs=` knob for `(•)`:* `D_ambiguous_width` blesses an opt-in knob but
   doesn't demand one, and `Checkbox`/`CheckboxGroup` both ship literals. Adding
   it here alone would create symmetry pressure for a third. Ship `(*)`/`( )`;
   add the knob to all three the day someone wants the bullet.
 - *A shared base with `CheckboxGroup`:* declined for the third time (see
-  `D-checkbox-group`). The two differ in exactly one line — `Set` membership vs
+  `D_checkbox_group`). The two differ in exactly one line — `Set` membership vs
   `==` — and the `cop` duplicate-rather-than-fold rule covers the rest.
 
 **Consequences.** Space on the already-selected row is a no-op, not a deselect:
@@ -1129,10 +1134,10 @@ selection and *both* rows render `(*)`, while two distinct items sharing a label
 stay independent (a row resolves to an item by index). The sampler pane reports
 value and cursor side by side, which is the cheapest way to see the split.
 
-## D-text-field-axes — `TextField`: two axes, horizontal scrolling, a logical cap (2026-07-31)
+## D_text_field_axes — `TextField`: two axes, horizontal scrolling, a logical cap (2026-07-31)
 
 **Status:** Accepted; `Component::TextField` rewritten 2026-07-31. Builds on
-`D-ambiguous-width` (which already asserted that "every rect, caret column and
+`D_ambiguous_width` (which already asserted that "every rect, caret column and
 clip derives from `StyledString#display_width`" — a claim `TextField` was quietly
 violating). Scoped to `TextField`; `TextArea` carries the same bug and is *not*
 fixed here.
@@ -1184,8 +1189,8 @@ half-painted.
 (`nil` by default), counted **in characters** — a wide glyph counts once — and it
 gates *typing only*: at the cap a printable key does nothing and is still
 consumed. It deliberately does not police `text=`, which stays authoritative as
-it is for `ComboBox#value` and `CheckboxGroup#value` (`D-combobox`,
-`D-checkbox-group`), so lowering the cap under an existing value leaves that
+it is for `ComboBox#value` and `CheckboxGroup#value` (`D_combobox`,
+`D_checkbox_group`), so lowering the cap under an existing value leaves that
 value intact instead of silently trimming it. A cap in *columns* was rejected: it
 would make the maximum text depend on which characters were typed, which is
 exactly the width-vs-length confusion this note removes.
@@ -1215,12 +1220,12 @@ or a long number is now reachable instead of rejected. `TextArea` is now the
 only component still conflating the axes — its wrap computation measures
 characters against a column width, so CJK prose overflows every row.
 
-## D-text-area-columns — `TextArea`: a cluster-iterating wrap over two axes (2026-07-31)
+## D_text_area_columns — `TextArea`: a cluster-iterating wrap over two axes (2026-07-31)
 
 **Status:** Accepted; `Component::TextArea` wrap rewritten 2026-07-31. The second
-half of `D-text-field-axes`, which fixed `TextField` and recorded this as open.
+half of `D_text_field_axes`, which fixed `TextField` and recorded this as open.
 Deliberately does **not** touch how the caret *steps* — that is
-`D-cluster-caret`.
+`D_cluster_caret`.
 
 **Context.** `compute_display_rows` filled each row by counting **characters**
 against `rect.width`, a **column** budget. So CJK prose wrapped at roughly twice
@@ -1285,17 +1290,17 @@ not just a bug fix, and it matches every editor.
   integer per row for a conversion on every mutation.
 
 **Consequences.** A row's `start` and `length` stay **character** counts, and
-`D-cluster-caret` kept them that way — boundary-locking the caret needed no
+`D_cluster_caret` kept them that way — boundary-locking the caret needed no
 change here at all, precisely because this wrap is already cluster-iterating and
 `chars_for_column` / `caret_to_display` already return boundary-aligned counts.
 The cluster-**width** question this entry left open was closed separately by
-`D-cluster-width`.
+`D_cluster_width`.
 
-## D-cluster-width — Emoji width policy `:rgi`; a cluster may exceed two columns (2026-07-31)
+## D_cluster_width — Emoji width policy `:rgi`; a cluster may exceed two columns (2026-07-31)
 
 **Status:** Accepted; implemented 2026-07-31. Completes the width story begun in
-`D-ambiguous-width` and continued through `D-text-field-axes` /
-`D-text-area-columns`, which fixed *where* widths were measured while this fixes
+`D_ambiguous_width` and continued through `D_text_field_axes` /
+`D_text_area_columns`, which fixed *where* widths were measured while this fixes
 *what a width is*.
 
 **Context.** Two independent bugs, both about the grapheme cluster as the unit a
@@ -1333,7 +1338,7 @@ cosmetics. `:rgi` is the only setting never wrong in the corrupting direction
 for a sequence it is exact when the terminal draws the parts and over-measures
 when the terminal combines them, and it treats VS16 emoji presentation as 2.
 
-Note this bets the *opposite* way from `D-ambiguous-width`, deliberately. That
+Note this bets the *opposite* way from `D_ambiguous_width`, deliberately. That
 note bets narrow because the glyphs at stake are Tuile's **own chrome** — box
 drawing, the scrollbar block — which the framework controls and needs at one
 column. Here the glyphs are **app content**, where the framework controls
@@ -1370,7 +1375,7 @@ rather than assumed.
   Rejected: it makes layout arithmetic non-reproducible across machines and
   makes the spec suite depend on whoever's `$TERM_PROGRAM` runs it — and Tuile's
   whole width strategy is one global answer with a small, enumerable inventory
-  (`D-ambiguous-width`). An app that needs its terminal's exact answer is better
+  (`D_ambiguous_width`). An app that needs its terminal's exact answer is better
   served by a future explicit override than by ambient detection.
 - **Clamp any cluster to 2 columns.** Would have avoided touching `put_char`,
   and is simply wrong for a non-RGI sequence the terminal really does draw
@@ -1384,15 +1389,15 @@ of its parts to 2, so any app that hard-coded the old number will disagree.
 return *fewer* columns than asked when a wide glyph straddles the boundary — it
 drops the glyph rather than halving it, as it already did for CJK. Unaffected: a
 cluster spanning two style spans takes the first span's style rather than being
-split. The caret stepped by character when this landed; `D-cluster-caret` fixed
+split. The caret stepped by character when this landed; `D_cluster_caret` fixed
 that separately.
 
 ---
 
-## D-screen-lifecycle — UI thread confinement, and three named screen states (2026-08-01)
+## D_screen_lifecycle — UI thread confinement, and three named screen states (2026-08-01)
 
 **Status:** Accepted; implemented 2026-08-01. First step of the tree-first
-sequencing (`D-tree-first`), and independent of the rest of it.
+sequencing (`D_tree_first`), and independent of the rest of it.
 
 **Context.** `Screen` carried a two-valued, unnamed state machine:
 `@pretend_ui_lock = true` in `initialize`, flipped to `false` on
@@ -1469,7 +1474,7 @@ raised `ENOTTY`, masking the real error.
 
 ---
 
-## D-tree-api — `@children` is authoritative; `add_child`/`remove_child` are the only path (2026-08-01)
+## D_tree_api — `@children` is authoritative; `add_child`/`remove_child` are the only path (2026-08-01)
 
 **Status:** Accepted and implemented 2026-08-01. No `children` override
 remains in `lib/`; the only `parent =` assignments left are the two inside
@@ -1478,7 +1483,7 @@ remains in `lib/`; the only `parent =` assignments left are the two inside
 **Context.** Five call sites used to hand-wire `child.parent = …` alongside
 their own child bookkeeping, each in its own order. That is where the
 transient tree inconsistency and the focus-repair ordering accident came
-from (`D-tree-first`), and it is what the attach/detach hooks would
+from (`D_tree_first`), and it is what the attach/detach hooks would
 have to fire *through*. Two shapes fix it, and they are not equivalent:
 
 - **A** — `Component` owns an `@children` array; `children` is a plain
@@ -1546,10 +1551,10 @@ for B.
 
 ---
 
-## D-attach-hooks — `on_attached` / `on_detached`: an edge trigger on the component (2026-08-01)
+## D_attach_hooks — `on_attached` / `on_detached`: an edge trigger on the component (2026-08-01)
 
 **Status:** Accepted and implemented 2026-08-01. Last step of the tree-first
-sequencing (`D-tree-first`); both `ideas/` notes it was designed in are retired.
+sequencing (`D_tree_first`); both `ideas/` notes it was designed in are retired.
 
 **Context.** Tuile had two thirds of a tree lifecycle: `attached?` (a computed
 predicate) and `on_child_removed` (a *container-side* notification used for
@@ -1606,7 +1611,7 @@ it must release in the mirror, because nothing else will.
   that goes on serving other sessions — so a missed `onDetach` there leaks into
   a *surviving* process and here it does not. That still holds, and it is why
   teardown-detach was never *urgent*; what overrode it is that `attached?`
-  became a type test (`D-tree-api`), so a tree rooted at a nilled `@pane` went
+  became a type test (`D_tree_api`), so a tree rooted at a nilled `@pane` went
   on claiming to be attached forever and touching it raised "Screen not
   initialized". Firing is also just cheaper than explaining that. So
   `Screen#close` now calls `ScreenPane#detach_all`.
@@ -1622,7 +1627,7 @@ it must release in the mirror, because nothing else will.
 - **A generic `Component#remove_all_children`** as the unmount primitive.
   Unsafe: a slot container calling it would empty `@children` while `#content`
   / `#footer` still pointed at detached components — exactly the desync
-  `D-tree-api` exists to prevent. Unmounting also has to clear the pane's own
+  `D_tree_api` exists to prevent. Unmounting also has to clear the pane's own
   slots, so it is not a generic tree operation. Named `detach_all` rather than
   `close` because `Popup#close` already means "remove *me* from the pane".
 
@@ -1637,16 +1642,16 @@ being detached (repair runs after), and must not inspect the ex-parent's
 bookkeeping. A raising hook propagates and leaves the tree undefined —
 durably so on the detach path, where the container's remaining work is skipped.
 Finally, hooks fire during `:idle` on the normal app path (a tree is assembled
-before `run_event_loop`), which `D-screen-lifecycle` made a decision rather
+before `run_event_loop`), which `D_screen_lifecycle` made a decision rather
 than an accident.
 
 ---
 
-## D-tree-first — `Screen` is the service, `ScreenPane` is the UI (2026-08-01)
+## D_tree_first — `Screen` is the service, `ScreenPane` is the UI (2026-08-01)
 
 **Status:** Accepted and implemented 2026-08-01, in five steps
-(`D-screen-lifecycle`, the one-axis `attached?`, `D-tree-api` in two parts,
-`D-attach-hooks`). The `ideas/` note it was designed in is retired.
+(`D_screen_lifecycle`, the one-axis `attached?`, `D_tree_api` in two parts,
+`D_attach_hooks`). The `ideas/` note it was designed in is retired.
 
 **Context.** Designing two no-op lifecycle hooks
 (`Component#on_attached` / `#on_detached`) took *ten* documented corner cases:
@@ -1670,8 +1675,8 @@ bookkeeping, each in its own order.
   `attached?` became `root.is_a?(ScreenPane)`: one axis, no `Screen`
   reference, so it never raises and a tree can be assembled with no screen in
   the process.
-- **The tree API is final** (`D-tree-api`), and `parent=` — reachable only
-  through it — is the sole lifecycle firing site (`D-attach-hooks`).
+- **The tree API is final** (`D_tree_api`), and `parent=` — reachable only
+  through it — is the sole lifecycle firing site (`D_attach_hooks`).
 
 Deleting the second axis deleted six edges outright rather than documenting
 them: the raise, the status-bar exception, the two-`@pane`-writes framing, the
@@ -1710,13 +1715,13 @@ means one screen, and the one-line change if that ever stops being true.
 
 ---
 
-## D-color-slots — A component color slot, not a new chrome token (2026-08-01)
+## D_color_slots — A component color slot, not a new chrome token (2026-08-01)
 
 **Status:** Accepted; first applied by `Component::ProgressBar#bar_color`
 (implemented 2026-08-02). Binds Slider and Badge when they land — the question
 was cross-component from the start, so it is settled once here rather than
-re-argued per widget. Builds on `D-bg-inherit` (accents-only theme, no global
-bg/fg token) and `D-theme-ref` (the live-resolved slot machinery this reuses).
+re-argued per widget. Builds on `D_bg_inherit` (accents-only theme, no global
+bg/fg token) and `D_theme_ref` (the live-resolved slot machinery this reuses).
 
 **Context.** {Theme} carries four chrome tokens — `active_bg_color`,
 `active_border_color`, `input_bg_color`, `hint_color` — and a component
@@ -1781,17 +1786,17 @@ member is additive, removing one is not.
   all. A `track_color` would have doubled the surface to weaken that.
 - **A slot's `Ref` is validated eagerly** (KeyError at assignment, as
   `bg_color=` does) and re-resolved at paint, never cached — same rules as
-  `D-theme-ref`, including riding the invalidate-everything pass on `theme=`.
-- **This licenses no global bg/fg token.** `D-bg-inherit` stands: a slot's
+  `D_theme_ref`, including riding the invalidate-everything pass on `theme=`.
+- **This licenses no global bg/fg token.** `D_bg_inherit` stands: a slot's
   `Ref` can only point at a color the theme *already* carries.
 
 ---
 
-## D-progress-bar — A value that is not a field; no text on the bar (2026-08-01)
+## D_progress_bar — A value that is not a field; no text on the bar (2026-08-01)
 
 **Status:** Accepted; `Component::ProgressBar` implemented 2026-08-02, demoed in
-the sampler. Color is `D-color-slots`; the glyph pair rides `D-ambiguous-width`;
-the ticker rides `D-attach-hooks`. What this entry owns is the *shape*.
+the sampler. Color is `D_color_slots`; the glyph pair rides `D_ambiguous_width`;
+the ticker rides `D_attach_hooks`. What this entry owns is the *shape*.
 
 **Context.** The first component with a `value` that is emphatically **not** an
 input: nothing focuses it, nothing types into it, and its number comes from the
@@ -1879,11 +1884,11 @@ paint over").
 
 ---
 
-## D-cluster-caret — The caret is boundary-locked; edits step by cluster (2026-08-02)
+## D_cluster_caret — The caret is boundary-locked; edits step by cluster (2026-08-02)
 
 **Status:** Accepted; implemented 2026-08-02 in `AbstractStringField`, so it
 landed on `TextField`, `PasswordField` and `TextArea` at once. Closes the gap
-`D-text-field-axes` / `D-text-area-columns` / `D-cluster-width` each recorded as
+`D_text_field_axes` / `D_text_area_columns` / `D_cluster_width` each recorded as
 open.
 
 **Context.** `@caret` indexed **codepoints** while the terminal draws **grapheme
@@ -1941,7 +1946,7 @@ makes the orphan bug unreachable.
   lines. Three concrete costs: (1) **it moves the axis, so every
   `caret = <something>.length` breaks silently** — five sites in `lib/` plus
   `examples/sampler.rb`'s `area.caret = start + command.length + 1`, all correct
-  for ASCII and wrong otherwise, which is the failure mode `D-text-field-axes`
+  for ASCII and wrong otherwise, which is the failure mode `D_text_field_axes`
   deleted, relocated from the framework to its callers; it then forced an open
   question about a loud rename migration purely to convert those silent breaks
   into `NoMethodError`s. (2) `max_text_length` would silently change meaning,
@@ -1961,7 +1966,7 @@ makes the orphan bug unreachable.
   repeated at every reader, and `caret=` already adjusts by clamping, so
   snapping there is not a new kind of surprise.
 - **Move `max_text_length` to counting clusters** alongside this. Deliberately
-  not bundled: it stays character-counting and stays `D-text-field-axes`'s
+  not bundled: it stays character-counting and stays `D_text_field_axes`'s
   decision. Now a knowing choice rather than an untouched default — a decomposed
   `é` burns 2 of 10, and a field at its cap refuses an accent on its last letter
   because `insert`'s check fires before the mark can merge.
@@ -1977,10 +1982,10 @@ an empty field, which is input validation, not an axis question.
 
 ---
 
-## D-float-field — `FloatField`: named for its Ruby type, and a deliberate copy of `IntegerField` (2026-08-07)
+## D_float_field — `FloatField`: named for its Ruby type, and a deliberate copy of `IntegerField` (2026-08-07)
 
 **Status:** Accepted; implemented 2026-08-07 (`Component::FloatField`). The
-`Float` half of `D-integer-field`'s "derived parse" case — same wrapper shape,
+`Float` half of `D_integer_field`'s "derived parse" case — same wrapper shape,
 same taxonomy slot, so only what *differs* is recorded here.
 
 **Context.** Vaadin calls this a *Number Field*; the survey in
@@ -2001,7 +2006,7 @@ category, not this field's value, and it would force the eventual sibling to be
 ~90% of their body (the `HasContent` shell, the `on_key` filter interceptor, the
 `fire_if_changed` guard) and differ in exactly the three places that matter: the
 filter, the parse, and the format. An `AbstractNumericField` with abstract
-`parse`/`format` hooks **is** the converter strategy `D-integer-field` kept out,
+`parse`/`format` hooks **is** the converter strategy `D_integer_field` kept out,
 reached through inheritance instead of a setter — and the `cop` rule is to
 duplicate rather than fold a shallow commonality into a base. The duplication is
 visible and boring; the base would be machinery.
@@ -2036,7 +2041,7 @@ spinner as `IntegerField`. A settable step is not free on a binary float:
 stepping by `0.1` accumulates `0.30000000000000004` straight into the visible
 buffer, so the knob would need a rounding policy (decimals? significant
 digits?), and rounding is formatting — a forms concern, parked with `min`/`max`
-in `D-integer-field`.
+in `D_integer_field`.
 
 **Alternatives rejected.**
 - *`BigDecimal` as the value type:* correct for money, but it needs the
@@ -2051,14 +2056,14 @@ in `D-integer-field`.
 
 ---
 
-## D-bigdecimal-field — `BigDecimalField`, and Tuile's first optional dependency (2026-08-07)
+## D_bigdecimal_field — `BigDecimalField`, and Tuile's first optional dependency (2026-08-07)
 
 **Status:** Accepted; implemented 2026-08-07 (`Component::BigDecimalField`).
-The third numeric field, so it inherits `D-float-field` wholesale (named for
+The third numeric field, so it inherits `D_float_field` wholesale (named for
 its Ruby value type, a deliberate copy rather than a shared base) — only the
 two things that are new are recorded here: exactness, and the packaging.
 
-**Context.** `D-float-field` closes with "the wrong field for money — hold that
+**Context.** `D_float_field` closes with "the wrong field for money — hold that
 as `Integer` cents"; this is the field that makes the honest answer available.
 `BigDecimal`, though, is not a language built-in: it was a *default* gem
 through Ruby 3.3 and became a **bundled** gem in 3.4, so from 3.4 on a Bundler
@@ -2124,16 +2129,16 @@ through the value seam: `"1.0"`→`"1.00"` fires nothing, because the two
   wanted, it belongs at the call site, where it is visible.
 - *A `scale=` / `decimals=` knob to pad the display:* it would have to rewrite
   the buffer under the caret while typing (`19.9` → `19.90` mid-edit), which
-  needs a blur/commit point a TUI lacks — the same reason `D-integer-field`
+  needs a blur/commit point a TUI lacks — the same reason `D_integer_field`
   gave for not normalizing.
-- *A settable `step=`:* `D-float-field` rejected it over binary-float noise,
+- *A settable `step=`:* `D_float_field` rejected it over binary-float noise,
   which genuinely doesn't apply here (`BigDecimal` steps exactly). Kept out
   anyway, so the three numeric fields stay one shape; this is the field to
   revisit first if the knob is ever wanted.
 
 ---
 
-## D-box-layouts — `Vertical` / `Horizontal`: declarative sugar with no `Auto` (2026-08-07)
+## D_box_layouts — `Vertical` / `Horizontal`: declarative sugar with no `Auto` (2026-08-07)
 
 **Status:** Accepted; implemented 2026-08-07 (`Component::Layout::Box`,
 `::Vertical`, `::Horizontal`, and the `Fixed` / `Percent` / `Expand` / `Insets`
@@ -2150,7 +2155,7 @@ PasswordField pane alone — renumbered by hand whenever a prompt gained a line)
 plus hand-rolled expansion (`[inner.height - 8, 2].max`) and hand-rolled cross
 clamps (`[inner.width, 30].min`). The cost was not that hand-rolling is
 impossible but that the code newcomers read to *learn* Tuile demonstrated the
-tedious version. The port took the sampler to 7 `Rect.new`.
+tedious version. The port took the sampler to a handful of `Rect.new` (5 today).
 
 **Decision — the vocabulary is `Fixed` / `Percent` / `Expand`, and there is no
 `Auto`.** Shrink-to-fit is the bottom-up `content_size` channel deleted in
@@ -2229,7 +2234,7 @@ override, so both class methods carry the guard (found by the spec, not by
 reading).
 
 **Decision — `Box` is a shared base, against the duplicate-don't-DRY rule.**
-`D-float-field` says duplicate rather than fold a *shallow* commonality into a
+`D_float_field` says duplicate rather than fold a *shallow* commonality into a
 base. This isn't shallow: the greedy pass is substantial and byte-for-byte
 identical except for which of `(left, top)` / `(width, height)` it reads, so
 `Box` parameterizes it behind two private hooks and `Vertical` / `Horizontal`
@@ -2269,11 +2274,10 @@ are ~10-line concretes. That is the sanctioned cohesive base
   is what a sidebar wants. Vaadin 8's ratio and CSS `flex-grow` both chose
   weights.
 - *`Min` / `Max` constraints:* deliberately not shipped, and the sampler shows
-  what that costs — its main split (`(width / 3).clamp(20, 40)`) and its two
-  sidebars (`min(16, width / 3)`) are caps on a *proportion*, unsayable in three
-  constraints, so they keep a rect-callback `Absolute`. That is the intended
-  division of labour: only the part needing arithmetic has any. Revisit only if
-  capped proportions turn out to be common.
+  what that costs — its two sidebars (`min(16, width / 3)`) are caps on a
+  *proportion*, unsayable in three constraints, so they keep a rect-callback
+  `Absolute`. That is the intended division of labour: only the part needing
+  arithmetic has any. Revisit only if capped proportions turn out to be common.
 - *`BorderLayout` / `BorderPane` / Textual's `dock:`:* nesting
   `Vertical(Fixed, Expand, Fixed)` covers it, and `ScreenPane` (content + status
   bar) already *is* one, hard-coded.
@@ -2309,7 +2313,7 @@ are ~10-line concretes. That is the sanctioned cohesive base
 
 ---
 
-## D-wrap-leading-space — An indent is content; no flag, and no hanging indent (2026-08-12)
+## D_wrap_leading_space — An indent is content; no flag, and no hanging indent (2026-08-12)
 
 **Status:** Accepted; implemented 2026-08-12 (`StyledString#wrap_one`). Fixes
 [issue #2](https://github.com/mvysny/tuile/issues/2). The continuation half —
@@ -2378,17 +2382,17 @@ so it is a no-op there, and the indented case is the only one with an opinion.
 
 ---
 
-## D-select — `Select`: the enum field, claiming no printable key but Space (2026-08-12)
+## D_select — `Select`: the enum field, claiming no printable key but Space (2026-08-12)
 
 **Status:** Accepted; `Component::Select` implemented 2026-08-12, demoed in the
-sampler. Builds on `D-has-value`, `D-combobox` (the chrome/value split and the
-resolve-don't-store-an-index rule, both adopted verbatim), `D-radio-group` (the
-cursor-is-chrome rule) and `D-ambiguous-width`.
+sampler. Builds on `D_has_value`, `D_combobox` (the chrome/value split and the
+resolve-don't-store-an-index rule, both adopted verbatim), `D_radio_group` (the
+cursor-is-chrome rule) and `D_ambiguous_width`.
 
 **Context.** A one-row closed-choice field: a face showing the selected item's
-label plus a `▾`, dropping open a `ListDropdown` of the options. `D-combobox`
+label plus a `▾`, dropping open a `ListDropdown` of the options. `D_combobox`
 deferred it once ("filterable first"), on the assumption that it needed the
-read-only-field axis `D-has-value` parked for the forms layer. That assumption
+read-only-field axis `D_has_value` parked for the forms layer. That assumption
 was an artifact of picturing a read-only `TextField` as the face; nothing gates
 this component.
 
@@ -2449,18 +2453,18 @@ dropdown as an overlay. Two consequences worth naming:
   having the child removes the wrinkle instead of documenting it.
 
 **Decision — promote `ComboBox#anchor` to `ListDropdown#anchor_to`.** Select needs
-byte-identical vertical geometry, and `D-float-field`'s duplicate-don't-DRY rule
+byte-identical vertical geometry, and `D_float_field`'s duplicate-don't-DRY rule
 **does not apply**: that licensed copying a *shell* around three genuine
 differences, whereas this is the same computation with zero differences, so a
 later fix to the flip rule would land in one copy and silently not the other —
 and the symptom appears only near a screen edge, which is invisible under test.
-The promotion threshold is the project's existing one (`D-color-slots`: "a
+The promotion threshold is the project's existing one (`D_color_slots`: "a
 *second* built-in needing the same thing"). Two rulings ride along:
 
 - **Width stays a caller-supplied parameter** (defaulting to the anchor's), so
   `ComboBox` keeps its lines-up-with-the-field policy and Select keeps its
   measured one, and `anchor_to` never measures content itself. Same shape as
-  `D-box-layouts`' "`align:` is legal only because the cross extent is
+  `D_box_layouts`' "`align:` is legal only because the cross extent is
   caller-supplied", and it keeps Select's measuring within the top-down re-grow
   rule: an optional, caller-side query feeding a rect the caller then assigns.
 - **Horizontally we slide, vertically we flip.** Covering the driver would hide
@@ -2508,7 +2512,7 @@ two callers that already know the answer.
   discover them one at a time by cycling, with no way to see the set or know how
   many there are. The dropdown is better at every item count, so the
   `ListDropdown` face is the only face, and the vocabulary does not grow a fourth
-  closed-choice widget (cf. `D-box-layouts`' "there is no `Auto`"). *Re-grow
+  closed-choice widget (cf. `D_box_layouts`' "there is no `Auto`"). *Re-grow
   rule:* if it returns it is a **face** on this component (a `dropdown: false`
   knob over the identical value seam), never a separate component, and it needs a
   real argument about visibility rather than a row-budget one.
@@ -2519,7 +2523,7 @@ two callers that already know the answer.
   one row itself.
 - *A shared base with `RadioGroup`* (`AbstractClosedChoiceField`): the ~15-line
   `items=` / `item_label=` / `label_for` shell is duplicated instead, per
-  `D-float-field`. The test is whether the commonality is a *shell around genuine
+  `D_float_field`. The test is whether the commonality is a *shell around genuine
   differences* or the *same computation* — `anchor_to` is the latter (extract), the
   items shell is the former (duplicate). The three differences a base would have
   to paper over with hooks: row rendering (`(*) label` glyphs vs. a bare label,
@@ -2527,7 +2531,7 @@ two callers that already know the answer.
   Space commits the row it's on, vs. the highlight *being* the pending selection),
   and where the rows live (always, in the component's own rect, vs. only while
   open, in a `Popup`'s). Three hooks over fifteen lines, reached through
-  inheritance, is the converter-strategy-by-inheritance shape `D-float-field`
+  inheritance, is the converter-strategy-by-inheritance shape `D_float_field`
   rejected — and it would couple two widgets that should stay free to diverge.
   This is the third copy of that shell, the same count `IntegerField` /
   `FloatField` / `BigDecimalField` reached; a *fourth* is when to re-argue it.
@@ -2545,7 +2549,7 @@ two callers that already know the answer.
   attempt is keystroke-driven, so it would flood a host's log on autorepeat, and
   an app may legitimately pass through item-less while loading. Enter/Space/Down
   are claimed either way — one rule, no branch. (An item-less Select is arguably
-  a *disabled* field, which touches the read-only/disabled axis `D-has-value`
+  a *disabled* field, which touches the read-only/disabled axis `D_has_value`
   parked for the forms layer. Not designed here, not foreclosed either.)
 - The dropdown is measured to the widest label plus `List`'s **two** row gutters
   (`pad_to_row` ellipsizes to `content_width - 2`, one leading and one trailing
@@ -2565,13 +2569,13 @@ two callers that already know the answer.
   `ComboBox#active=` leans on, and filtering / row rendering / the commit action
   really do vary.
 
-## D-list-items — `List` takes items + a renderer, rendered lazily (2026-08-14)
+## D_list_items — `List` takes items + a renderer, rendered lazily (2026-08-14)
 
 **Status:** Accepted; implemented 2026-08-14, with the five composers folded onto
-it in the same series. Builds on `D-has-value` (typed, not stringly),
-`D-combobox` (resolve an index, never store one), `D-float-field` (duplicate
+it in the same series. Builds on `D_has_value` (typed, not stringly),
+`D_combobox` (resolve an index, never store one), `D_float_field` (duplicate
 rather than fold a shallow commonality) and the top-down layout rule
-(`D-box-layouts`). Delivers the first half of the "typed items + data provider on
+(`D_box_layouts`). Delivers the first half of the "typed items + data provider on
 `List`" item that gated List Box, Grid and Virtual List.
 
 **Context.** `List` took pre-rendered rows: `lines=` stored `Array<StyledString>`
@@ -2583,7 +2587,7 @@ seam:
   because the framework handed back a string.
 - Four components (`ComboBox`, `Select`, `RadioGroup`, `CheckboxGroup`) kept a
   private copy of the `@items` / `@item_label` / `label_for` / `rebuild_rows`
-  shell. `D-select` set the trigger for re-arguing a shared base at the *fourth*
+  shell. `D_select` set the trigger for re-arguing a shared base at the *fourth*
   copy; this is it.
 
 **Decision — externalize rendering on the generic component.** `List` holds
@@ -2594,7 +2598,7 @@ follows elsewhere — a domain component takes data, a generic one takes strateg
 
 **Not a shared base class.** The alternative reading of four duplicated shells is
 "extract `AbstractItemsComponent`". That is exactly the `parse`/`format`-hook base
-`D-float-field` rejected, one level up: it would need a render hook, a
+`D_float_field` rejected, one level up: it would need a render hook, a
 commit-gesture hook and a where-do-rows-live hook to span a dropdown driver and a
 row-per-item group. The duplication was a symptom of a missing *seam*, not of a
 missing *ancestor*, and adding the seam deleted the duplication that actually
@@ -2704,12 +2708,12 @@ having `value=` rebuild every row is the O(n) pass this decision just deleted.
   rendering — is the shape that makes the bug unsayable, and the PTY test now
   presses Enter.
 
-## D-scroll-nomenclature — `row` is the grid, `line` is `\n`, `items` are domain objects (2026-08-14)
+## D_scroll_nomenclature — `row` is the grid, `line` is `\n`, `items` are domain objects (2026-08-14)
 
-**Status:** Accepted; implemented 2026-08-14. Builds on `D-list-items` (which
-made the item vocabulary real), `D-text-area-columns` and `D-text-field-axes`
+**Status:** Accepted; implemented 2026-08-14. Builds on `D_list_items` (which
+made the item vocabulary real), `D_text_area_columns` and `D_text_field_axes`
 (which named the index-vs-column axes inside the inputs) and
-`D-ambiguous-width` (whose `display_width` is the column authority).
+`D_ambiguous_width` (whose `display_width` is the column authority).
 
 **Context.** Three scrolling components had grown three vocabularies for the
 same four concepts — a content unit, a wrapped unit, a viewport-relative row,
@@ -2787,7 +2791,7 @@ which this scheme makes unwriteable — `line` is never a coordinate.
   stop using row-words loosely. Naming no row is not an exception to "row
   everywhere".
 - **`List#items` → `List#rows`,** which a List item arguably is. Rejected:
-  `items` is where `cop` wants the domain-object noun (`D-list-items` had just
+  `items` is where `cop` wants the domain-object noun (`D_list_items` had just
   landed it), and it is the word the enum widgets above `List` already use.
 - **`scroll_top`** (CSS's `scrollTop`, shorter). Rejected for `scroll_top_row`:
   it names no unit, and `list.scroll_top` reads as an imperative — *scroll to
@@ -2815,7 +2819,7 @@ which this scheme makes unwriteable — `line` is never a coordinate.
   wrong.
 - **`row_count` was reserved here, then created separately.** Making it a public
   reader was held to be a behavioural addition needing its own argument; that
-  argument is `D-text-area-rows`, which granted it on `TextArea` only. The point
+  argument is `D_text_area_rows`, which granted it on `TextArea` only. The point
   this entry settled — that the *name* is already taken, so the addition need not
   re-litigate its spelling — held.
 - **`CHANGELOG.md` was not swept.** Its 0.4.0 entry announcing the `set_line` /
@@ -2823,10 +2827,10 @@ which this scheme makes unwriteable — `line` is never a coordinate.
   and describes what shipped *then*, so retro-editing it would make a released
   migration note reference a method that release did not have.
 
-## D-text-area-rows — `TextArea#caret_row` / `#row_count`, not a hook and not the wrap (2026-08-15)
+## D_text_area_rows — `TextArea#caret_row` / `#row_count`, not a hook and not the wrap (2026-08-15)
 
 **Status:** Accepted; implemented 2026-08-15. Grants the reader
-`D-scroll-nomenclature` reserved the name for. Answers
+`D_scroll_nomenclature` reserved the name for. Answers
 [#3](https://github.com/mvysny/tuile/issues/3).
 
 **Context.** Shell-style prompt-history recall in a `TextArea`: Up recalls the
@@ -2866,7 +2870,7 @@ the auto-growing prompt strip — the case the name was reserved for — uses
   0.9.0 layout re-grow rule explicitly sanctions the opposite — capability
   returning as "an *optional, read-only, caller-side query* … never as an
   automatic channel the framework consults" — which is also why
-  `D-scroll-nomenclature` rejected a general `Component` scroll seam. It serves
+  `D_scroll_nomenclature` rejected a general `Component` scroll seam. It serves
   one question, in one direction, at one moment, where the readers also serve the
   prompt strip, a "row 3/7" readout and a caller-drawn scrollbar. And it needs a
   subclass, where the readers serve `on_key` too. In COP terms it is neither a
@@ -2898,7 +2902,7 @@ the auto-growing prompt strip — the case the name was reserved for — uses
   responsible for keeping them consistent. **(4) It flips a written invariant for
   no argued caller** — AGENTS.md holds the class private "until a second caller
   actually exists", and nobody has asked for `row_text` / `index_at` from
-  outside. Forwarders grow on demand at one line each; `D-float-field`'s
+  outside. Forwarders grow on demand at one line each; `D_float_field`'s
   temperament ("a fourth copy is when to re-argue it") applies.
 - **A `wrapped_text` method documented "do not store".** Same staleness, renamed.
 - **A validity token on `WrappedText`,** so a holder can detect a stale snapshot.
@@ -2922,15 +2926,17 @@ the auto-growing prompt strip — the case the name was reserved for — uses
   which is the whole reason the object stays private. Specs pin that both track a
   text change and a width change.
 
-## D-text-view-scroll-verbs — `TextView#scroll_half_page_up` / `#scroll_half_page_down` (2026-08-15)
+## D_text_view_scroll_verbs — `TextView#scroll_half_page_up` / `#scroll_half_page_down` (2026-08-15)
 
-**Status:** Accepted; implemented 2026-08-15.
+**Status:** Accepted; implemented 2026-08-15. Amended 2026-08-23: the
+`active?` guard this was originally argued *from* turned out to be dead code
+(see the correction at the end) — the decision stands on its other grounds.
 
 **Context.** A chat TUI keeps focus in the input field beneath its transcript,
-so the transcript's own scroll keys never fire: {Component::TextView#handle_key}
-opens with `return false unless active?`. The host wants PageUp/PageDown at the
-*prompt* to page the *view*, half a screen at a time so the reader keeps an
-overlap while output streams in. `TextView` already knows how to do exactly
+so the transcript's own scroll keys never fire: dispatch delivers a key along
+the focus chain only, and the view is not on it. The host wants PageUp/PageDown
+at the *prompt* to page the *view*, half a screen at a time so the reader keeps
+an overlap while output streams in. `TextView` already knows how to do exactly
 that — `Ctrl+U` / `Ctrl+D` have scrolled by half a viewport since the scroll
 ladder landed — but every clamped primitive behind those bindings
 (`move_scroll_top_row_by`, `move_scroll_top_row_to`, `viewport_rows`,
@@ -2952,8 +2958,11 @@ it never learns the row count, never clamps, and never touches focus.
   that wants it, where the two spellings drift. `TERMINOLOGY.md` also pins
   `viewport_rows` private on purpose — `rect.height` is its public form.
 - **Let the host forward a synthetic key** (`view.handle_key(Keys::CTRL_U)`).
-  Dead on arrival — the `active?` guard rejects it, which is the whole problem —
-  and a keystroke aimed at an unfocused widget is a lie about where focus is.
+  A keystroke aimed at an unfocused widget is a lie about where focus is: the
+  host's question is "scroll this view", and spelling it as a key makes the
+  view's key bindings part of its API — rename `Ctrl+U` and the caller breaks.
+  (At the time this was also *dead on arrival*, the guard rejecting it; that
+  guard is gone and the forward would now work. It is still the wrong spelling.)
 - **App-side arithmetic on the existing public `scroll_top_row=`.** It raises
   below `0` and is deliberately *not* clamped above, so a caller who overshoots
   the last row leaves `at_bottom?` false and silently kills `auto_scroll`
@@ -2962,7 +2971,7 @@ it never learns the row count, never clamps, and never touches focus.
   "Page" should page, it would break `Ctrl+U`/`Ctrl+D`'s reason to exist, and it
   fixes nothing anyway: an unfocused view still sees no keys.
 - **Ship the whole ladder as verbs** (full page, top, bottom, by-row). No caller
-  yet; `D-text-area-rows`'s temperament applies — a future caller argues its own
+  yet; `D_text_area_rows`'s temperament applies — a future caller argues its own
   case, and these two settle the spelling for the rest.
 
 **Consequences.**
@@ -2979,10 +2988,26 @@ it never learns the row count, never clamps, and never touches focus.
   paging back to the last row re-arms it. The host gets read-while-streaming for
   free and has nothing to wire.
 
-## D-notification — One corner toast, N messages, one ticker draining them (2026-08-17)
+**The correction (2026-08-23).** `TextView#handle_key`'s opening
+`return false unless active?` was **vestigial**, and this entry took it for a
+live constraint. It was a leaf backstop for the one place the pre-0.8 framework
+over-delivered (`ScreenPane` forwarding to `content` unconditionally); e1777fe
+centralized dispatch and dropped the same guard from `TextInput`, `List` and
+`Button` — but `text_view.rb`, three weeks old at the time, was missed. It could
+never fire once removed from that context: `bubble_key` walks `Screen#focused`
+upward and `focused=` marks that chain `active`, and a `TextView` is a leaf, so
+the only chain position it can hold is `focused` itself. The stale
+`return true if super` above the `case` went with it — `Component#handle_key`
+has collapsed to `false` since the same commit. Both lines are deleted; the
+widget now obeys the framework-wide rule (AGENTS.md, book ch5) that a
+`handle_key` acts on the key alone. The *visible* change is that hand-feeding a
+key to an unfocused view now scrolls it, which is what every other widget in the
+gem already did (`examples/sampler.rb`'s unfocused `List` is the house idiom).
+
+## D_notification — One corner toast, N messages, one ticker draining them (2026-08-17)
 
 **Status:** Accepted and implemented, `Component::Notification`. Builds on
-`D-attach-hooks` (the synced-from-an-invariant ticker), `D-color-slots` (the
+`D_attach_hooks` (the synced-from-an-invariant ticker), `D_color_slots` (the
 per-message color), and Tier 1 of the component survey. Book ch7 "Notifications"
 is the user-facing half; the rdoc owns the per-symbol contract. What this entry
 owns is *why each choice*, and the alternatives that looked right first.
@@ -3074,7 +3099,7 @@ makes "at most one" true.
 - Corollary for a future factory: `self.show` calls bare `new`, never
   `Notification.new`, so a subclass's `show` builds the subclass.
 - This widget is what surfaced `Popup.self.open` as a subclass trap (it had to be
-  privatized here too, until the factory was deleted outright — `D-popup-open`).
+  privatized here too, until the factory was deleted outright — `D_popup_open`).
 
 ### The singleton lives in the popups stack, never in a class ivar
 
@@ -3192,11 +3217,11 @@ second *kind* of anchoring that would unlock it (per the component survey), but
 `Notification` ships its own `reposition` first so the extraction is judged with
 two real implementations rather than one and a guess.
 
-## D-popup-open — No class-level `Popup.open` factory; `#open` returns `self` (2026-08-17)
+## D_popup_open — No class-level `Popup.open` factory; `#open` returns `self` (2026-08-17)
 
 **Status:** Accepted and implemented; `Component::Popup.open` **removed**, and
 `Popup#open` now returns `self`. Surfaced while building
-{Tuile::Component::Notification} (`D-notification`), which had to privatize the
+{Tuile::Component::Notification} (`D_notification`), which had to privatize the
 inherited factory to stop it undermining a private constructor.
 
 **Context.** `Popup.open(content:, modal:, size:)` was one-line sugar for
@@ -3236,3 +3261,2162 @@ needs the popup *before* mounting it in order to wire `on_pick`.
 own arguments, and wraps the popup rather than *being* one — none of them is an
 inherited factory, so the trap does not apply. `popup_spec` asserts that neither
 `Popup` nor `ListDropdown` responds to `open` at the class level.
+
+## D_bracketed_paste — A paste is its own event, not a burst of keys (2026-08-23)
+
+**Status:** Accepted and implemented in `Keys` (`BRACKETED_PASTE_ON`,
+`PASTE_START`, `read_paste`, `normalize_paste`), `EventQueue::PasteEvent`,
+`Screen#run_event_loop(bracketed_paste:)`, `ScreenPane#handle_paste`,
+`Component#handle_paste`, `AbstractStringField#handle_paste`, and
+`FakeScreen#paste`. Reported as
+[issue #4](https://github.com/mvysny/tuile/issues/4).
+
+**Context — the two bytes are the same byte.** Pressing Return in raw mode sends
+`\r`. Pasting into a terminal that has *not* been told the app can tell a paste
+apart also sends `\r` for every clipboard line break: xterm, VTE and tmux all
+rewrite the selection's `\n` on the way out, deliberately, so that a paste looks
+exactly like typing (tmux's `paste-buffer -r` exists to opt out of it). So a
+{Tuile::Component::TextArea} subclass that rebinds ENTER to submit — the
+chat-prompt shape — submitted **once per pasted line**, and the first line was
+gone before the second arrived.
+
+Nothing downstream can repair that. By the time `handle_key("\r")` runs, "the
+user pressed Enter" and "the clipboard held a line break" are the same event.
+The only downstream lever is inter-keystroke timing, which `D_select` already
+rejected for type-ahead on exactly this ground: a terminal degrades that signal
+(bytes in one read burst merge into a single key) and a paste has no gaps at all.
+The information exists only at the layer that talks to the terminal, which is
+Tuile's.
+
+**Decision — drive DEC private mode 2004, on by default.** `run_event_loop`
+prints `\e[?2004h` alongside the mode-2031 notify and `\e[?2004l` in the same
+`ensure`, and takes `bracketed_paste: false` to opt out, mirroring
+`capture_mouse:`. Terminals that don't know the mode ignore the sequence, so
+there is no capability probe and nothing to detect — which is what makes
+defaulting it *on* safe rather than a gamble. The off switch exists for the same
+reason `capture_mouse: false` does: a terminal that mishandles the mode, and a
+one-flag escape beats a fork of the loop.
+
+**Decision — the payload is read raw, not through `Keys.getkey`.** `getkey`
+returns `\e[200~` cleanly (its 5-byte tail gulp fits the marker exactly), but the
+*content* must not go back through it: a pasted `\e` would send it gulping five
+bytes of clipboard as an escape tail and surfacing them as phantom keypresses —
+the failure the `\e[M` and `\e[?` drains already exist to prevent. So
+`Keys.read_paste` reads **one byte at a time** to the `\e[201~` terminator.
+One byte at a time, and not a chunked read, because a chunk would over-read past
+the terminator and swallow whatever the user typed behind the paste; there is
+deliberately no pushback buffer in `Keys` to make chunking safe. A paste is
+human-scale and arrives once, so the syscall count is not worth a second
+mechanism.
+
+**Decision — a `PasteEvent`, and it never touches the key ladder.** The key
+thread posts one event carrying the whole payload; `Screen#event_loop` routes it
+to `handle_paste` down the focus chain, with the same modal scoping as a key and
+no other rung. *Rejected: reusing `KeyEvent` with a flag*, which would put a
+`pasted?` predicate on the ladder and re-create the runtime gate `D_key_dispatch`
+deleted — every `handle_key` would have to remember to check it, and the ones
+that forgot would be exactly today's bug. *Rejected: replaying an unhandled paste
+as individual keys.* It reads like graceful degradation and is the ambiguity
+walking back in through the fallback: a component that declines a paste would
+still get eight ENTERs. Unhandled text is dropped.
+
+**Decision — the field inserts it as one mutation.**
+`AbstractStringField#handle_paste` inserts at the caret in a single `text=`, so
+`on_change` fires once for the paste rather than once per character. That is what
+lets a submit-on-Enter subclass need *no* paste code at all — it keeps
+`handle_key` for the typed ENTER and inherits paste-inserts-text — and it
+incidentally retires an O(n) re-render and, for a slash-command overlay, an O(n)
+re-filter.
+
+**Where each layer sanitizes, and why the line is there.** `Keys.normalize_paste`
+fixes only what is a *terminal* artifact: `\r`/`\r\n` → `\n` (terminals disagree
+about which they send inside the brackets — readline carries its own `\r`→`\n`
+pass for precisely that reason, so this cannot be left to the caller), and an
+invalid-UTF-8 scrub so a pasted binary file cannot make a downstream
+grapheme-cluster walk raise. Control characters are *content* and survive that
+layer. What a **text buffer** may hold is the field's call:
+`AbstractStringField#preprocess_paste` drops the C0 controls (a raw `\e` or `\t`
+reaching {Tuile::Buffer} would move the real cursor mid-frame), keeps `\n`, and
+turns a tab into one space rather than inventing a tab width;
+`TextField#preprocess_paste` narrows further — newlines to spaces, since a
+one-row field holds no line break, and a trim to `max_text_length` rather than a
+rejection, because that is what typing the same characters would have done. An
+app wanting tab *expansion* or a `[Pasted 230 lines]` placeholder overrides
+`handle_paste`, which is the seam that exists for it.
+
+**Testing is three layers, because no one of them covers the others.**
+`FakeScreen#paste` (normalize + dispatch) is the unit door and starts one layer
+above the terminal; the sampler's *Paste* pane is the visual demo; and one PTY
+example in `spec/examples/sampler_spec.rb` is the only place mode 2004, the
+marker recognition and `read_paste` run for real. That PTY test writes the whole
+`\e[200~…\e[201~` sequence as **one burst**, which is the one place AGENTS.md's
+pace-the-keys rule is deliberately inverted: a real paste *is* a gapless burst,
+and the payload is drained raw, so nothing in it can be mistaken for a key. Its
+assertions read newly painted log rows rather than the counter row, because the
+buffer flushes the minimal diff — `rows in draft: 1` becoming `…: 3` puts one
+character on the wire, not the phrase.
+
+**Corrected while here.** `TextArea`'s rdoc claimed a pasted line break arrived
+as `\n` and a typed one as `\r`, which is backwards and read as though multi-line
+paste already worked. Accepting {Keys::CTRL_J} is still right, but its
+justification is now the honest one: that is the byte a *typed* Ctrl+J sends.
+
+## D_repaint_cascade — the repaint cascade skips the clear, never the invalidate (2026-08-23)
+
+**Status:** Accepted and implemented in {Tuile::Component#repaint}. Found while
+building {Tuile::Component::TabSheet}, but the bug predates it and was already
+visible in three shipped sampler panes.
+
+**Context — the symptom.** Focus the sampler's *TabSheet* pane and press Tab to
+put focus on the strip: the pane below it vanishes. It is still there — Tab once
+more and it comes back — so nothing was detached; the cells were simply blanked
+and never repainted. The same fault, less dramatically, blanked four rows of the
+*Checkbox*, *CheckboxGroup* and *RadioGroup* panes whenever focus moved into
+them. A sweep comparing each pane's incremental repaint against a
+repaint-everything baseline is what found the other three.
+
+**The mechanism, in one chain.** A focus change invalidates every component
+whose `active?` flipped — i.e. the whole new focus chain. One of those is a
+`Layout::Vertical(spacing: 1)`, whose children leave gaps, so the default
+{Tuile::Component#repaint} runs `clear_background` over **its whole rect** —
+which is every descendant's cells, not just the gaps — and then re-invalidates
+its *direct children*. That notice then has to travel the rest of the way down,
+and it didn't: the old default opened with
+
+    return if children.any? && children_tile_rect?
+
+so a container whose children tile it perfectly painted nothing **and
+re-invalidated nothing**. A `TabSheet` (strip on row 0, pane below, exactly
+tiling) is such a container, and so is a `Layout` whose slot happens to fit its
+children. The cascade dead-ended there, the grandchildren never learned their
+cells had been wiped, and the blank stayed until some unrelated event invalidated
+them again. Nothing in the code says "this must forward", and no test went red —
+the invalidation set and the buffer were both self-consistent.
+
+**Decision.** Make the *clear* conditional and the *invalidate* unconditional:
+
+    clear_background unless children.any? && children_tile_rect?
+    children.each { |c| screen.invalidate(c) }
+
+A container that paints nothing of its own can only redraw its area *through* its
+children, so being invalidated has to mean invalidating them. The tiling test
+keeps doing the one job it is good for — deciding whether there is a gap worth
+blanking, which is what `D_progress_bar`'s "never blank a cell you are about to
+paint over" cares about.
+
+**Why the extra invalidation is not a cost.** It is a repaint of a subtree that
+was about to be wrong, and it reaches the terminal only if it changes something:
+`Buffer::Cell#set` flips the dirty flag on a real content change alone, so
+repainting identical glyphs emits nothing. The wire stays minimal; only CPU
+moves, and only on the frames where an ancestor cleared.
+
+**Roads not taken.**
+
+- **Clear only the gaps instead of the whole rect.** Strictly better in
+  principle — no descendant's cells would be destroyed, so no cascade would be
+  needed at all — but it means real rect-subtraction geometry (n children, holes,
+  overlap) in the hottest path in the framework, to replace one `fill`. The
+  cascade fix is three lines and needs no new geometry. Revisit only if clearing
+  ever shows up in a profile.
+- **Fix it in `TabSheet` alone** (invalidate the strip and pane from its own
+  `repaint`). Rejected on evidence: the sweep proves three other panes already
+  had the bug, so the fault is the framework default, not the new component. A
+  local fix would have left the trap armed for the next container that happens to
+  tile.
+- **Make the clearing container invalidate the whole subtree** (`on_tree`) rather
+  than its direct children. Same end state by a blunter route, and it moves the
+  knowledge of "who might have been clobbered" into the clearing parent, where
+  the tree below it is none of its business. Each container forwarding one hop is
+  the local rule that composes.
+
+## D_tabs — `Tabs` / `TabSheet`: a strip, and a strip that swaps panes (2026-08-23)
+
+**Status:** Accepted; `Component::Tabs` (with `Tabs::Tab`) and
+`Component::TabSheet` implemented 2026-08-23, demoed in the sampler, taught in
+book ch7 ("Switching between views"). Brainstormed in `ideas/tabs.md`, now
+retired. Leans on `D_has_value` (the seam it declines), `D_progress_bar` (the
+precedent for a selection kept *out* of that seam), `D_list_items` (items vs.
+identities), `D_select` (claim the minimum), `D_ambiguous_width` (the separator
+glyph), `D_tree_api` (the slot-swap recipe) and `D_attach_hooks` (what
+detachment fires).
+
+**Context.** Several views, one visible at a time, and a one-row strip of
+captions to pick between them. Two components, because the strip is useful
+alone — Vaadin documents that case explicitly ("content switching without Tab
+Sheet"), and an app whose strip lives structurally elsewhere on the screen
+needs it: `Tabs` is the selector, `TabSheet` is the selector plus the pane that
+goes with it.
+
+**Decision — neither is `HasValue`, because a selection is not a value.** The
+test that decides it, and it generalizes: **would a form save it?** A
+`RadioGroup`'s selection *is* the datum being edited, so it is a value; a tab's
+selection is where the user is looking — nothing saves it, nothing validates it,
+and a forms layer iterating fields must never find it. `D_progress_bar` made the
+same call one step further out (a `value` that is a read-only report), and
+`List` has held a cursor and an `on_item_chosen` without being a field since it
+existed. External corroboration: **Vaadin's `Tabs` is not a field either** — it
+fires `SelectedChangeEvent`, exposes `setSelectedTab`/`setSelectedIndex`, and is
+grouped with Accordion and Details rather than with the fields. The cost is that
+`Tabs` gets no `empty?` / `clear` / `on_value_change` and no free `focusable?`,
+so it declares `focusable?` and `tab_stop?` itself, the way `Checkbox` and
+`Select` do. Someone will eventually ask for `tabs.value`; the answer is
+`selected` / `selected_index`, and this paragraph is why.
+
+**Decision — `on_tab_selected` reports that the selection *changed*, not that
+the user pressed something.** Arrows, a click, `selected=` / `selected_index=`,
+the autoselect of the first `add_tab`, and the re-selection that follows removing
+the selected tab all fire it; re-selecting the tab already selected fires
+nothing. Removing the *last* tab fires `(nil, nil)`, both arguments nil. The
+alternative — notify only on user gestures — would make `Tabs` the one component
+where an app must re-derive the selection after a removal, and the empty case is
+exactly where a listener most needs to hear from the strip: an app that renders
+from the callback has to be told to render *nothing*, or the departed tab's
+content sits on screen with no tab pointing at it. One implementation
+consequence worth keeping: the notification decision cannot be made by comparing
+indices, because removing the selected middle tab of three leaves the index at 1
+with a *different* tab under it. `apply_selection` therefore takes the
+previously-selected tab as an argument.
+
+**Decision — hiding a pane means *detaching* it; Tuile grows no visibility
+flag.** `TabSheet` keeps only the selected tab's pane in the tree. The
+alternative — n+1 children, unselected panes hidden by an empty rect — looks
+cheaper and is not, because the empty rect is a *paint* convention that gates
+nothing else. Five leaks, all silent: `cycle_focus` collects tab stops by tree
+walk, so every field in every hidden pane stays in the Tab cycle;
+`first_tab_stop_or_root` and `Layout#on_focus` cascade focus *into* hidden
+subtrees; `Screen` parks the hardware cursor at `focused.cursor_position`, so a
+hidden `TextField` puts the terminal cursor in the middle of the visible pane;
+`keyboard_hint` advertises the hidden widget in the status bar; and key delivery
+bubbles through it because it is on the focus chain. Only mouse hit-testing is
+safe. So option B needs a real seam gating at least four places plus a ruling on
+whether `Box` / `Absolute` skip invisible children when dividing space — a
+framework-wide change in the focus system, to buy one component what detachment
+already gives. Note the prior art: every framework that keeps hidden panes
+mounted (Textual, FTXUI) has a display/visibility flag in its *core* — Textual's
+`ContentSwitcher` is one `display` toggle. Tuile's honest options were detach or
+invent that flag. **Re-grow rule:** `Component#visible?` comes back only when a
+*second* consumer appears (a pane that must stay live while hidden, an app
+wanting hidden-but-laid-out widgets), and only argued as a focus-and-paint gate
+with an explicit ruling on layout arithmetic — never as a paint-time flag
+smuggled in under one component. AGENTS.md carries the one-line invariant.
+
+**Decision — one tab stop for the whole strip, and arrows activate
+immediately.** Three arguments against a component per tab, in order of force:
+"exactly one stop per widget" (`D_has_value`), and n tabs would mean n Tab
+presses before the content is reachable; making the *Tab key* walk between
+*tabs* is the one thing the key ladder forbids by construction (Tab is claimed
+above everything and means "leave this widget"), so it would read as a feature
+and be a semantic inversion; and no prior art does it, including the frameworks
+where individual tabs are widgets (Textual's `Tab`s are children of a focusable
+`Tabs` and are not focus stops).
+
+Activation is immediate — Textual, Terminal.Gui, FTXUI, Windows tab controls and
+the ARIA "automatic activation" pattern all agree, Vaadin being the lone
+counterexample with a manual variant motivated by expensive panels and
+screen-reader semantics a TTY doesn't have. The deciding reason is narrower than
+the prior art, though: **auto-activation means only one thing is ever
+highlighted.** Manual activation needs two states on one row — the selection and
+the roamed-to tab — and therefore two visual channels to separate them, on a
+strip that spends both on the selection alone (below). `RadioGroup` could afford
+that split vertically because each row has a glyph column of its own
+(`D_radio_group`); a one-row strip cannot, and two highlights side by side read
+as noise rather than as two kinds of state. Auto-activation deletes the
+distinction instead of styling it, and every code path — paint, hit test,
+callback — has one index to consult. Consequence, and it runs the opposite way
+to the brainstorm's guess: **lazy panes inherit this rather than reopening it.**
+Arrowing across five lazy tabs builds five panes; a sheet that can't afford
+that owes its own answer (a cheap placeholder, or building on a settle delay),
+not a return to Enter-to-activate.
+
+**Decision — the strip claims LEFT / RIGHT and the mouse, and nothing else.**
+`D_select`'s contract restated: Enter and Space have nothing to do once arrows
+activate, and declining them keeps a form's default button and the app's keys
+alive. UP / DOWN are declined so a future arrow-navigating layout can move focus
+*out* of the strip on the axis the strip doesn't use. HOME / END are declined
+too — Terminal.Gui binds them on its tab row, but a key no widget claims stays
+available app-wide, which is worth more than a shortcut for a jump that is two
+Left presses away in the 3–5 tab normal case; an app that wants it assigns
+`selected_index`. Edges clamp and consume, no wrap (as `List` does): the
+arrow-nav rule about declining at the edge is about *focus motion*, and this is
+selection, with the vertical axis already the way out.
+
+**Decision — bold marks the selected tab; it is not strip chrome.** The selected
+caption is bold *always*, and additionally sits on `Theme#active_bg_color` while
+the strip is on the focus chain; unselected captions are regular weight. Two
+channels, no new theme token. Bold is the one that survives an unfocused strip,
+which matters because the strip is the map of where you are in the app — unlike
+a `List` cursor, which is a transient pointer and has
+`show_cursor_when_inactive` for exactly this reason. **Bolding every caption is
+the tempting "fix" and it is wrong**: it spends the only unfocused-visible
+channel, leaving selection to the focus-gated background alone, so an unfocused
+strip would show no selection at all. Neither escape works — `input_bg_color` is
+the only other bg token and it means "resting input well" (`Select` uses it for
+exactly that), and dimming the *unselected* captions instead collides with the
+dim a *disabled* tab wants (see Deferred below), which would leave unselected
+and disabled indistinguishable.
+Rejected alternatives: bracketing the label (`[Payment]`) shifts every later
+segment by two columns whenever the selection moves, making hit-test geometry
+depend on the selection; an underline is `▁`, a fresh Ambiguous glyph. This
+ruling also needed one new primitive — `StyledString#with_bold`, since nothing
+in the gem had used bold and a caption is a `StyledString` that may carry its own
+colors.
+
+**Decision — the separator is `│`, the glyph `Window` paints its borders with,
+not ASCII `|`.** This inverts `D_ambiguous_width`'s "a new component defaults to
+ASCII when the pretty glyph is Ambiguous", and the inversion is the point: that
+rule exists to keep the Ambiguous inventory small and enumerable, and `│` is
+already *in* the inventory — `window.rb` paints it on every window, and nothing
+in the gem is designed to survive it measuring 2. Reusing a glyph the framework
+has already bet on adds nothing to the audit list, and a strip inside a window
+lines up with the border around it. `separator=` remains, now as the opt-in for
+ASCII. A *fresh* Ambiguous glyph still defaults to ASCII.
+
+**Decision — segment geometry: the padding is part of the segment, the separator
+column is chrome.** A segment is `" " + caption + " "`, segments joined by one
+separator column. Every segment has the same shape including the first and last
+(no trimmed outer padding, so no edge case in the arithmetic); the highlight
+covers the padding, because one that stopped at the glyphs would read as a
+ragged smear; and a click on a padding column selects that tab, while the
+separator column selects nothing — same rule as the blank tail past `extent`,
+which focuses without selecting (`D_boolean_fields`). One private `segments`
+method is the sole source of that arithmetic, read by *both* the paint and the
+hit test, and derived from the captions on each call rather than recorded during
+the last paint — so a hit test is correct before the first paint and after a
+caption change.
+
+**Decision — a narrow strip scrolls to keep the selection whole in view**
+(2026-08-24; v1 clipped, and this replaces that ruling before either strip
+shipped). One private `left_column` — the strip column painted in the rect's
+leftmost cell, the name `TextField` uses — read by the paint, the hit test,
+`extent` and (on `MenuBar`) the segment rect a panel anchors to, so there is
+still exactly **one** source of segment arithmetic; two would let a click land on
+the tab beside the one drawn under it. One idempotent `adjust_left_column` is its
+sole writer, called from every mutation site (`ProgressBar#sync_ticker`'s shape,
+not a nudge per site), which is what makes the offset `0` in every situation the
+clipping version handled, scroll back on its own when the rect grows or the
+captions shrink, and never need a scroll-back branch in any mutator. `MenuBar`
+funnels its three highlight writers through one private `highlight=` for the same
+reason, and gets a guarantee out of it: the highlighted segment is on screen
+*before* `Cascade` anchors a panel to it.
+
+Rejected — *segment-aligned scrolling* (the offset always a segment start): it
+buys clean edges and needs no glyph snapping, but wastes up to a segment of width
+at the right edge, and a strip this narrow is exactly where columns are scarce.
+Rejected — *reserved cue columns*: reserving two columns makes the window width a
+function of the scroll state that is computed from it, which is `D_select`'s
+`:auto`-scrollbar circularity, and shifts the whole strip sideways when a caption
+is edited. The cues are **overlaid** on the edge columns instead, keeping the
+style of the cell they cover so one landing on the selected segment doesn't punch
+a hole in its highlight, and they are ASCII `<` / `>` — `‹ ›` are Ambiguous-width
+(`D_ambiguous_width`), and a `cue_glyphs=` knob with no caller is a knob to argue
+about later. They stay chrome, not buttons: a click on a cue falls through to the
+half-visible segment under it, which selects it and reveals it — the direction the
+cue pointed anyway — where a clickable cue would need the column to hit-test
+differently from what it paints. And *free scrolling* (a wheel moving the window
+without moving the selection) is deliberately absent: the next sync would yank the
+view back to the selection, so supporting it means a second "user scrolled, stop
+following" state with a resume rule — `List#auto_scroll`'s machinery, for a
+one-row widget.
+
+**The trap this design steps over.** `StyledString#slice` *drops* a grapheme
+cluster straddling the window edge rather than half-painting it, so an offset
+landing mid-cluster returns a row one column short and shifts everything past the
+hole one column left — paint and hit test then disagree, silently, only for wide
+glyphs. So the offset is snapped *forward* to a cluster boundary, as
+`TextField#snap_to_glyph_start` does; forward is the safe direction, giving up at
+most one column of the segment left of the window and never of the one being
+revealed. A caption wider than the whole rect can't be shown whole at all: its
+head wins, being the half that identifies it.
+
+**Decision — `Tabs` owns mutable `Tab` handles; it does *not* get the
+`items=` / `item_label=` / `label_for` shell.** `add_tab("First")` mints and
+returns a `Tabs::Tab`, Vaadin-style. The test that separates the two is sharper
+than "items feel wrong": **an item is an element of a collection someone else
+owns** — assignment is whole-collection, and an item carries no per-element
+state, the renderer deriving everything from the object each paint (which is why
+`D_list_items` *removed* the appenders). **A tab is identity plus per-element
+mutable state**, minted by the widget and living as long as it, and re-assigning
+the whole set — the operation an items API is built around — is precisely what a
+strip must never offer: it would destroy tab identity and with it `TabSheet`'s
+pane mapping. Two corollaries make the ruling durable: the unbuilt half of
+`D_list_items` is a *data provider* behind `items`, and a provider cannot own
+per-tab state, so `HasItems` would arrive carrying a promise Tabs must refuse
+(paging tabs is meaningless — a million tabs is not a UI); and the growth path
+here is per-element *attributes* (hidden, disabled, closeable), which items have
+no notion of. So the `HasItems` question is closed for `Tabs`; it survives only
+for `ComboBox` / `Select` / `RadioGroup`, where the shell genuinely is three
+copies of one thing.
+
+The synthesis worth keeping: **the `Tab` object is what keeps those attributes
+from becoming framework seams.** A hidden tab is a skipped segment, not
+`Component#visible?`; a disabled tab is painted dim and skipped when arrowing,
+not a framework enabled/disabled seam; a closeable tab is an `x` in the segment.
+None touches `Component`.
+
+`Tab` is a small mutable object owned by the strip — not a frozen value type (it
+has settable attributes) and not a `Component` (it never paints itself; a
+component that never paints is a confusing new category). The contract is copied
+wholesale from `TextView::Region`: `private_class_method :new`, handed out by the
+owner, mutators invalidating the owner through a back-pointer, and **a removed
+handle raising on every mutator and on every reader that consults the strip** — a
+stale `Tab` is the same footgun as a stale `Region`. As there, the locally-held
+`caption` stays readable (so an error message can name it) and `remove` is an
+idempotent no-op. The back-pointer also closes the caption-refresh question:
+`Tab#caption=` invalidates the strip, so there is no `refresh_rows`-style
+question to answer.
+
+**Decision — no `Tab#data`.** A tab carries a caption and its own display
+attributes, nothing of the app's. The rejected slot would have let
+`on_tab_selected` hand back a domain object, and it isn't needed: the pane
+component owns its data (COP's "a component does everything its one purpose
+needs", so the pane *is* the handle), or a future binder does — the tab is on
+neither path. Anything genuinely per-tab and app-owned lives in the `TabSheet`
+or the app component that built it, keyed the way `TabSheet` keys its panes.
+This is what keeps `Tab` from becoming the items API this entry just refused.
+
+**Decision — `Tab` hand-rolls `caption` / `caption=` rather than including
+`HasCaption`.** The six duplicated lines look like exactly what a mixin is for,
+and the reason they aren't is the mixin's actual payoff: `HasCaption` earns its
+place as a **test-locator seam** — a locator walks the component tree matching
+`is_a?(HasCaption)` plus a caption compare, with no hardcoded class list. A
+`Tab` is not a `Component`, so it appears in no tree walk and that payoff is
+structurally unreachable; and there is exactly one `Tab` class, forever, so the
+"no hardcoded class list" benefit has nothing to range over either. Including it
+would be DRY-only, which is the bar the seam argument sets. The lookup debt is
+paid on the strip instead: **`Tabs#tabs`** returns the tab array (read-only by
+convention, like `Component#children`), so a test finds a tab through the widget
+that owns it — `tabs.find { |t| t.caption.to_s == "Payment" }`. That
+reader was needed anyway, since `TabSheet` keys panes by identity and nothing
+else can enumerate.
+
+**Decision — `TabSheet` holds two children, not n+1, and is not
+`HasContent`.** `children == [strip, pane]` with the strip pinned at index 0, so
+pre-order traversal yields the browser's strip-then-pane Tab order for free.
+`HasContent` stays out even though the swap looks like a content slot, for three
+concrete reasons: `content=` would become public API meaning "the visible pane",
+which is misleading (the pane is *derived* from the selection, not assignable);
+`HasContent#handle_mouse` forwards only into `content`, so the strip would never
+see a click; and `HasContent#on_focus` forwards focus into the content, which is
+the behavior this design rejects (switching a tab must not move focus into the
+new pane — browser and Vaadin behavior). What *is* reused is the slot-swap
+recipe `D_tree_api` specifies for `Window`: detach without notifying, rewire,
+then `on_child_removed` last, so the focus repair cascades into the *new*
+occupant. `TabSheet` overrides that hook to land focus on **the strip** rather
+than on itself, which is not focusable; the other candidate (the new pane's
+first tab stop) loses because the user's last action was a tab switch.
+
+Panes live in an identity-keyed `Tab => Component` map on the sheet. Licence:
+`Box`'s per-child constraint map, which AGENTS.md permits because it is "a
+per-child *attribute* map, not a second copy of ordering" — the strip's tab array
+stays the sole ordering authority. Rejected: a `component` slot on `Tabs::Tab`
+(the strip would then know about panes, which is the split this whole design
+rests on), and `TabSheet::Tab < Tabs::Tab` behind a protected factory hook (a
+framework hook existing for exactly one subclass, handing the minting decision
+to the subclass while `Tabs` still owns the array).
+
+Two implementation rulings the map earned. **One idempotent `sync_pane` is the
+sole writer of the visible pane**, deriving it from `strip.selected` on every
+call — which is what lets `add_tab` register a pane *after* the strip has already
+autoselected its tab, with no suspend-the-listener dance; the event-driven
+alternative has an ordering problem on the very first tab. And the map's keys are
+kept honest by an invariant rather than by one code path:
+`forget_removed_tabs` drops every entry whose tab is detached, because
+`Tabs::Tab#remove` reaches the strip without passing through
+`TabSheet#remove_tab` — which stranded the entry, pinned the pane against GC, and
+made `add_tab` reject that pane as still in use. Rejected there: an
+`on_tab_removed` listener on `Tabs` for the sheet to subscribe to — the tidier
+data flow, but new app-facing API whose only consumer is internal.
+
+Named `add_tab(caption, pane)`, not `add`: `Layout#add(component)` is the house
+`add`, and the explicit verb stops the two reading alike — the same reason
+`Tabs#add_tab` isn't `add`. (The brainstorm sketched `sheet.add`; its own
+argument overruled it.)
+
+**Decision — no framework key switches tabs from *inside* a pane, and no
+`Keys::CTRL_PAGE_UP` / `CTRL_PAGE_DOWN` constants are added.** Not v1, not
+later. Four reasons, the first decisive: it is **a global shortcut in disguise**
+— "one key, anywhere in the app, meaning switch tab" is app policy, and Tuile
+already has two homes for app policy (the rung-2 registry and an ancestor's
+`handle_key`), so shipping it as component behavior smuggles an app-level binding
+into a widget. Nested sheets make it ambiguous *and* the failure is silent: the
+bubble delivers to the innermost `TabSheet` first, so an inner sheet swallows the
+key and the outer one becomes unreachable by keyboard with nothing on screen
+explaining why. Vaadin apps have never needed it — the strip plus Tab is enough.
+And the editors that do have it use their own scheme, which is the argument for
+leaving the binding to the app: no choice Tuile made here would match the app's
+other keys. What Tuile owes instead is the *verbs*: `select_next` /
+`select_previous` are public (they exist for Left/Right anyway), so an app that
+wants the habit writes two lines and owns both the key and the "which sheet"
+question that sank the framework version.
+
+**Deferred, and why each lands additively.** v1 is captions, selection, mouse,
+keys and the pane swap. Nothing below is blocked, which is the payoff of the
+`Tab`-object ruling — each is a `Tab` attribute plus a branch in paint and in
+arrowing, needing no framework seam:
+
+- **Hidden tabs** — skip the segment when painting and when arrowing. Pane
+  hiding is already detachment, so this needs no `Component#visible?`.
+- **Disabled tabs** — paint dim, skip when arrowing, never select. A disabled
+  *tab* is not a component, so no enabled/disabled seam is needed. (A disabled
+  *pane* would be; still out of scope.)
+- **Closeable tabs** — an `x` in the segment, hit-tested, removing the tab.
+  Nobody else in the gem needs it and the glyph is ASCII-cheap.
+- **Lazy panes** — `add_tab("Reports") { build_reports }`, built on first
+  selection (Vaadin does it with an attach listener). Free to add: the swap has
+  one call site. Inherits auto-activation, per the activation ruling above.
+- **Clickable cues, and free scrolling** — see the scrolling decision.
+
+**Alternatives rejected** (beyond those argued inline). *Vertical orientation:*
+out of scope — Vaadin doesn't allow it in a TabSheet either, and a vertical strip
+is a `List` with a renderer (the Side Nav shape). *A border around the strip:*
+compose with `Window`; the Turbo Vision / Terminal.Gui look, with tabs notched
+into the top border, would couple `Tabs` to `Window` chrome. *Prefix/suffix
+slots* for icons and badges: unnecessary — a caption is a `StyledString`, so
+`Open [24]` is just text.
+
+**Consequences.** Selection is view state, so nothing in a forms layer will ever
+enumerate a strip. Hiding a component means detaching it, framework-wide, and
+`TabSheet` is the worked example — which also means a pane's `on_attached` /
+`on_detached` fire on every switch, and a pane cannot own a resource that must
+outlive its visibility. A tab is a handle an app holds, so tab identity is stable
+across caption edits and reorderings of nothing else. And a starved strip stays
+wholly reachable, at the cost of a scroll offset that every future paint-time or
+hit-test change has to keep threading through one place.
+
+## D_menu_bar — `MenuBar`: a focused strip driving a cascade of `ListDropdown`s (2026-08-24)
+
+**Status:** Accepted; v1 (`Component::MenuBar` with `MenuBar::Item` and the
+private `MenuBar::Cascade`) implemented 2026-08-24, demoed in the sampler, taught
+in book ch7 ("Menus"); v2 (mnemonics) the same day. Designed in a since-retired
+`ideas/menu-bar.md`, whose prior-art survey (Vaadin 25.2, Turbo Vision,
+Terminal.Gui, notcurses, MC, and the frameworks that have no menu) this entry
+only summarizes.
+
+**Update 2026-08-24: a narrow bar scrolls**, on `D_tabs`' scrolling decision,
+which both strips implement identically (one private `left_column`, one
+`adjust_left_column` as its sole writer, ASCII cues overlaid on the edge
+columns). `MenuBar`'s share of it: one private `highlight=` funnels the arrow,
+mnemonic and click paths, so a segment is on screen before `Cascade` anchors to
+it, and `rect=` still *closes* the cascade rather than re-anchoring it.
+
+**Context.** `ideas/new-components.md` listed Menu Bar as blocked on extracting a
+`Popover` from `ListDropdown#anchor_to`. It isn't: the widget needs a *second
+placement*, not a second kind of overlay.
+
+**Decision.** Focus never leaves the bar. The strip is the single tab stop, and
+the open menus are non-modal `ListDropdown`s mounted on the `ScreenPane` — owned
+by the bar, parented by nobody — so every key arrives at `MenuBar#handle_key`,
+which offers it to a `Cascade` first. That is `Select`'s architecture
+(`D_select`) extended to N levels, which is why **nothing in the key-dispatch
+ladder changes** and why the whole widget is additive: two new placement helpers
+on `ListDropdown`, one callback pass-through, and no change to `Popup`,
+`ScreenPane` or `Component`.
+
+The keyboard map is copied from Vaadin's, which is also the ARIA menubar pattern
+and what every TUI lineage surveyed does — Left/Right along the strip,
+Down/Enter/Space to open, Up/Down inside, Right/Enter to drill, Left to go back,
+ESC to close one level, and Left-at-the-top / Right-on-a-leaf stepping to the
+neighbouring menu. There was nothing to invent, and inventing would have been the
+error.
+
+**Alternatives rejected.**
+
+- **A single drill-down frame** — one panel that re-renders as you descend
+  (Terminal.Gui ships this as `UseSubMenusSingleFrame`). It needs no
+  `anchor_beside` and no stack at all, and was rejected because it loses the
+  "where am I in the hierarchy" readout that is the cascade's entire point — and
+  because it cannot be the default with a cascade bolted on later: the cascade is
+  the harder mechanism, and building it second means building it against a shape
+  that assumed one panel.
+- **A modal level-0 popup**, which would give real modality — keys scoped, clicks
+  outside blocked. Rejected on a mechanical fact, not a preference:
+  `ScreenPane#add_popup` **centers** every modal popup and focuses it, so an
+  anchored modal is impossible without changing `ScreenPane`. It would also
+  invert ownership, moving key handling off the bar and into the popup.
+- **Focusable panels**, focus descending as you drill. Rejected: AGENTS.md's
+  non-modal-overlay traps say a focus-taking non-modal overlay lands focus
+  outside the key scope and kills *every* keystroke until Tab recovers. This
+  would be that bug once per level.
+- **Extracting `Popover` now.** The roadmap's own trigger ("the second *kind* of
+  anchoring") arguably fires here, but both callers still wrap a `List`, so a
+  `Popover < Popup` would move code without a second kind of *content*. The
+  trigger is the first non-`List` content wanting anchoring (Tooltip, a
+  date-picker grid). It originally had a second half — a third placement method
+  on `ListDropdown`, from `ContextMenu`'s `anchor_at(point)` — which went dormant
+  when that widget was iced (`D_no_context_menu`).
+- **A command-code bus** (Turbo Vision's `cmOpen` + `handleEvent`) instead of
+  per-item callables. Rejected: Ruby has closures, and Vaadin, Terminal.Gui and
+  ratatui's `tui-menu` all landed on per-item listeners.
+- **`item.submenu` as a separate object** (Vaadin's `getSubMenu()`). It exists
+  because a Vaadin `MenuItem` is a DOM component; a Tuile item is a handle, so
+  `item.add_item` is one hop shorter and makes depth fall out for free.
+- **`Component::MenuItem` as a top-level constant.** Considered and reverted the
+  same day: promoting it was priced against a breaking rename once `ContextMenu`
+  names the type, and that price is zero, since no release ships in between. With
+  the cost gone the house default (`Tabs::Tab`, `List::Cursor`) wins, and
+  `MenuBar` gets to settle as one coherent component before unification is argued
+  against a second implementation rather than a guess about one. **Now settled
+  rather than deferred:** icing `ContextMenu` removed the counterparty, so
+  `MenuBar::Item` is simply the name. A revival after 0.13.0 ships pays a
+  **Breaking:** changelog line for the rename, or keeps `MenuBar::Item` as an
+  alias — cheap, and only paid if it happens, which beats paying it now for a
+  widget that may never exist.
+- **A `HasMenuItems` mixin** (Vaadin's shared `MenuBar` / `ContextMenu` /
+  `SubMenu` interface). Not needed yet, and the shape keeps it cheap: `MenuBar`
+  delegates `add_item` / `items` to a captionless root `Item`, so the method
+  exists exactly *once* and a future sharing exercise starts from one
+  implementation rather than two that drifted.
+- **Separators (`add_separator`).** Looks free, isn't: a `List` has no
+  unselectable row, so the cursor would land on a separator and Enter would
+  activate nothing. It needs a `Cursor` that hops non-selectable positions, which
+  is a `List` decision, not this one.
+
+**Deliberately not like `Tabs`.** The strip reuses `Tabs`' *hit testing* — an
+`extent`, one private `segments` method feeding both paint and click — and
+deliberately not its *look*: no separator column, no bold, and no highlight while
+unfocused. Both are one-row caption strips with one highlighted segment, so
+looking alike would leave a reader working out which control they are seeing. Two
+of the three divergences are forced anyway: bold is `Tabs`' *persistence* channel
+(the selection must survive focus moving on) and a menu bar has nothing to
+persist. The segments arithmetic is the second copy of that trio; per AGENTS.md's
+duplicate-a-shallow-shell rule a third caption strip is when to argue for
+extraction.
+
+**Consequences a contributor would trip over.**
+
+- **Activation is uniform.** Children win over a listener; a leaf closes the
+  cascade *before* firing (so an action that opens a dialog doesn't paint it under
+  a menu, as in `Select#commit`); and an item with **neither** children nor a
+  listener is legal and inert. An item that looks live but does nothing is the
+  app's error to fix, not the framework's to raise on.
+- **Stepping highlights; only Enter, Space or a click presses.** Left/Right
+  moving to a neighbouring *top-level button* (a listener, no menu) closes the
+  cascade and highlights it — it does not fire it, or walking the strip would
+  trigger every button on the bar, each one behind a menu still standing over its
+  output. Every path that *does* fire a top-level listener closes the cascade
+  first, matching `Cascade`'s own leaf activation.
+- **A resize closes the menu**, from `MenuBar#rect=` — see the AGENTS.md
+  non-modal-overlay trap for why that is the legal answer here rather than a
+  `reposition` override. Only a *changed* rect closes it, since `Layout::Box`
+  re-assigns an equal rect on any child mutation.
+- **So does detaching**, from `on_detached`: the panels are the pane's children,
+  not the bar's, so nothing else would take them down.
+- **An open menu swallows keys; a closed strip does not.** The one deliberate
+  divergence from `D_select`'s claim-the-minimum rule, and the honest reading of
+  what a menu is — an app key firing behind a visible panel is worse than a dead
+  keystroke.
+- **A click outside an open cascade is not blocked**, because non-modal overlays
+  block nothing — but it does *dismiss*. The framework-level fix this entry
+  called for (and declined to invent here) shipped as `D_outside_click`: the
+  pane closes every popup a left click missed, and `Cascade` reconciles its level
+  stack from each panel's `Popup#on_close`. The click itself still reaches
+  whatever is beneath.
+- **`Cascade` is provisional.** It is split from the strip on cohesion, not reuse
+  — otherwise `MenuBar` would both paint captions and manage an overlay stack —
+  and the test for keeping it is *the size of the interface `MenuBar` needs*: at
+  `open_below` / `handle_key` / `close` / `open?` it is a boundary; if it grows
+  accessors that expose the level stack, the "class" was only ever a seam and it
+  folds back in.
+- **Widths are measured per level, caller-side**, third repeat of the `D_select`
+  pattern (`anchor_to` and `anchor_beside` measure nothing). The submenu arrows
+  right-align against the level's *widest label*, a number the cascade already
+  has, so they line up without asking the `List` how wide it ended up.
+- **The `▸` is Neutral, not Ambiguous** — verified, like `Select`'s `▾`. The
+  obvious `▶` / `▼` are Ambiguous and would have needed an ASCII opt-in under
+  `D_ambiguous_width`.
+
+**Mnemonics (v2), and why they are legal.** `add_item(caption, mnemonic: "f")`
+at *every* depth. AGENTS.md deleted `Component#key_shortcut` and the capture
+phase that scanned a scope subtree, and forbids reintroducing them — but its
+re-grow rule sanctions exactly this: *sugar over an ancestor's `handle_key`,
+never a dispatch phase and never a gate*. A focused `MenuBar` consulting its own
+item tree inside its own rung-3 `handle_key` is unregistered, unscanned and
+invisible to every other component. This is the first thing a reviewer will
+(correctly) flag, hence the paragraph.
+
+The rule is **one live set, no fallback**: the top-level items while the cascade
+is closed, the deepest open panel's items while it is open, nothing else ever
+consulted. Cross-level collision is therefore *structurally impossible* rather
+than tie-broken — `File > Export` and top-level `Edit` may both bind `e`, and
+with File open there is nothing to arbitrate — and `f`,`q` for File > Quit falls
+out with no chord, buffer or timeout. A duplicate *within one sibling set* raises
+at `add_item`, which is the only scope where two mnemonics can race. A miss
+swallows rather than falling back to a shallower level: a mistyped letter must
+not tear down the open menu and open another, and Left/Right and ESC are the
+routes to a different menu. All of this is what Windows/GTK/Qt do; macOS is the
+only lineage without menu mnemonics, for the historical reason that it never had
+an Alt-activates-the-menubar model.
+
+Four consequences worth recording, each a road that looked open:
+
+- **The match is hoisted above the cascade delegation.** v1's cascade swallows
+  every unrecognized key while open, so a letter would never reach the strip
+  otherwise. It is guarded to a single printable non-space character so Enter,
+  Space, the arrows, ESC and `MOVE_KEYS` keep their v1 path.
+- **The cue is `Item#cued_caption`, computed once at construction.** There are
+  two paint sites (the strip's segments, the cascade's row renderer) and
+  `StyledString#slice` counts **columns** while a caption search yields a
+  **character** index, so the conversion lives in exactly one place. Safe to
+  precompute — unlike a theme value, it has no live input. Cues are **always
+  drawn**, focused or not: Tuile has no Alt to reveal them with, so the choice is
+  binary and discoverability wins.
+- **The bell is tied strictly to the swallow**, and guarded by `Keys.printable?`
+  at the swallow site. Unguarded it would ring at HOME, function keys and the
+  five-byte junk `Keys.getkey` returns for an unknown escape sequence. No bell
+  while the strip is *closed* — a bubbled key is not a miss — and none for a
+  matched-but-inert item or a clamped arrow, or "beep when nothing happened"
+  would grow into an audit of every no-op path.
+- **`List#select(index)` was the one real gap.** The cascade must move a panel's
+  highlight to the matched row *before* drilling, or a submenu anchors beside
+  whatever row the cursor was on — and a row scrolled out of view has no rect to
+  anchor against at all. `List` could move its cursor by key, by mouse and by
+  search, but not by index; that hole is independent of menus.
+
+**Type-ahead search is deliberately not built.** "Type `s` in an open menu to
+jump to the first item containing s" is nearly free — `List#select_next` already
+does substring, case-insensitive, cursor-ordered-with-wrap search — and that is
+the trap: it competes with explicit mnemonics for the same keystroke, so it owes
+a precedence rule *and* a ruling on whether a unique match fires or merely
+highlights. A separate feature, for a later session.
+
+**Deferred, each additive:** checkable and disabled items, global-shortcut
+activation (which needs `Keys` to grow function keys first — and this is the
+deferral that costs something, since with no Alt the only way to *reach* the bar
+is Tab, which is what separates `Alt+F, X` from a Tab-hunt), removal and
+reordering, dynamically computed items, open-on-hover
+(needs mouse motion — Tuile runs X10 mode 1000, press-only), and Vaadin's
+collapse-into-an-overflow-menu.
+
+**Update 2026-08-24: `ContextMenu` is iced indefinitely** — designed, priced and
+declined the same day, in `D_no_context_menu`. It would have reused `Cascade` and
+`Item` verbatim, which is why the two consequences above are worded the way they
+are: the nested `Item` name is *settled* rather than deferred, and the `Popover`
+extraction trigger keeps only its "first non-`List` content" half.
+
+## D_outside_click — An outside click dismisses a popup, by flag not by notice (2026-08-24)
+
+**Status:** Decided and implemented 2026-08-24. Designed in a since-retired
+`ideas/outside-click-dismiss.md`, itself split out of the declined `ContextMenu`
+(`D_no_context_menu`), so this entry is the whole record. Supersedes the wart
+`D_menu_bar` recorded without fixing.
+
+**Context.** Whether an open overlay closed when you clicked elsewhere depended
+on what you happened to click *on*. A click on a focusable widget moved focus,
+and losing focus is what closed `Select`'s dropdown and `MenuBar`'s cascade — so
+it worked, by accident. A click on decoration (a `Label`, a `Window` border, a
+gap between fields, the status bar row) did nothing at all, and the overlay
+stayed open over content it no longer belonged to. Three customers felt it:
+`Select`, `MenuBar`'s whole cascade, and the sampler's slash-menu demo.
+
+It could not be fixed inside the widgets. `ScreenPane#handle_mouse` routes a
+click to the topmost popup containing it, else the tiled content, else (with a
+modal open) nobody — so a click that misses every popup is never reported to the
+open overlay, no driver can poll for it, and nothing below can forward it. A
+`ScreenPane` change or nothing.
+
+**Decision.** `Component::Popup#close_on_outside_click?` (default `true`, modal
+or not), read by `ScreenPane#handle_mouse`: a left click that misses an open
+popup closes it. Beside it, `Popup#on_close`, a driver-facing callback fired from
+`on_detached`.
+
+**Why a flag and not `on_outside_click(event)`.** The rejected alternative was
+notice-shaped: every missed popup gets the event, default no-op, with a
+driver-facing proc beside it. It works, and it is more expressive. It was
+rejected because it hands a `MouseEvent` to a component that is *not* on the
+chain the event was delivered to — structurally the same second delivery this
+project already rejects for a `Screen`-level click broadcast, just with a shorter
+subscriber list. Under the flag, `ScreenPane` never delivers anything twice: it
+closes popups that asked in advance to be closed. **The popup receives a fate,
+not an event**, and "a click is delivered exactly once, down one chain" stays
+literally true.
+
+The price is expressiveness: the popup answers with a stored `true`/`false`, not
+with an opinion about the click. Paid once, by `ComboBox` — its field is tiled,
+so clicking your own input to reposition the caret closes the list you are
+filtering. Transient, because `TextField#on_change` is wired to `refill`, which
+reopens it on the next keystroke, and Vaadin's ComboBox behaves the same way. If
+per-click nuance is ever genuinely needed, widen the reader to
+`close_on_outside_click?(event)` — a pure widening, no migration — rather than
+reaching for a notice or a veto.
+
+**The ordering rule, both halves load-bearing.** Snapshot the open popups
+*before* routing, close the opted-in misses *after*.
+
+- *Snapshot before*, or a popup the delivered click **opened** is in the set and
+  dismisses itself instantly — every `Select` would be unopenable by mouse.
+- *Close after*, or a widget toggling its own overlay from a click on its face
+  sees a shut overlay and **reopens** it — a `Select`'s dropdown could then never
+  be dismissed by clicking the Select.
+
+Both are specced, and both mutations also break *pre-existing* `Select` specs.
+The snapshot is a fresh array for a third reason: a handler may close further
+popups, and `@popups` must not be mutated mid-iteration.
+
+**"Outside" spans the owner chain, and stacking order plays no part.**
+`Popup#owner` names the component an overlay is *part of* (`nil` = an overlay in
+its own right). A click keeps the popup it hit *and* every popup that one
+belongs to, transitively; everything else dismissable closes. The owner is any
+`Component` — a driver hands its dropdown `self` — and the pane resolves it to
+the enclosing popup at click time, a `Popup` resolving to itself.
+
+Two bugs forced this, both found by clicking rather than by reasoning, and both
+after the naive "closed if it missed my rect" rule had shipped:
+
+- **A cascade panel is beside its parent, not inside it.** Drilling by mouse
+  (File → Open Recent → Archive) dismissed every shallower panel, so the File
+  menu vanished the moment you clicked into its own submenu.
+- **A dropdown routinely hangs past its dialog's border.** A `ComboBox` or
+  `Select` on a dialog's lower rows drops a panel outside the dialog's rect, so
+  clicking a row dismissed the dialog. The most common form layout there is.
+
+Neither is reachable by a widget-local fix: the panels and the dialog are
+different popups with no way to speak for each other.
+
+**Two rejected rules, and why order is the wrong axis.** *Dismiss the popups
+stacked above the one you clicked* (standard light-dismiss layering) fixes both
+bugs with no new API, and was rejected because `@popups` is insertion order and
+Tuile has no click-to-raise: the same click would produce different outcomes
+depending on which overlay opened first. *A click on any overlay dismisses
+nothing* also fixes both, needs no API at all, and was rejected because it
+declares unrelated overlays related — it leaves a dropdown open when you click
+the dialog beneath it, and stops two window-like overlays from dismissing each
+other, which is exactly what they should do.
+
+Order is only ever the *shadow* of ownership: a child overlay cannot exist
+before its host, so it is always later in the stack. Reading the shadow works
+for related popups and is meaningless for unrelated ones, which is why the
+relationship is declared instead.
+
+Consequences kept: every dismissable popup closes, not just the topmost — a
+cascade must vanish whole on one background click, not peel one panel per click
+— and two *unrelated* stacked modals both close on one outside click, where
+Vaadin's curtain would close only the top. Arguably Vaadin-consistent anyway:
+the Flow Dialog docs say closing a modal Dialog also closes the dialogs opened
+after it.
+
+**The cost, stated plainly.** `owner` is a declaration you can forget, and
+forgetting it silently reproduces the two bugs above. It is the third entry in
+AGENTS.md's non-modal-overlay traps for that reason. Three sites wire it today:
+`ComboBox` and `Select` hand their dropdown `self` at construction (not per
+open, so there is nothing to forget on reopen), and `Cascade#push` chains each
+panel to the one it dropped out of. Level 0 owns nothing on purpose — a click on
+a dialog hosting the bar *should* close the whole menu and keep the dialog. A
+mis-wired cycle terminates rather than hanging, guarded by the walk.
+
+**Why `on_close` hangs off `on_detached`, never `#close`.** A popup leaves the
+screen three ways — `Popup#close`, a direct `Screen#remove_popup`, and
+`Screen#close` → `detach_all`. Hang the proc off `#close` and two of those vanish
+silently, which is the desync the mechanism exists to kill, reintroduced one
+level up. `parent=` is already the sole firing site for the lifecycle hooks, so a
+proc over `on_detached` keeps that true and makes the notice unconditional. The
+subclass trap that follows: `Notification#on_detached` already existed and now
+calls `super`.
+
+`MenuBar::Cascade` is the worked example and the reason the callback exists. It
+keeps `@levels` as the sole authority on depth, so a panel closing behind its
+back would leave `depth` / `deepest` / `highlighted` all lying. It wires an
+identity-keyed, idempotent delete — idempotent because the same notice also
+arrives from its own `truncate` (which has already popped the entry) and from
+teardown, in no guaranteed order. That is the shape the house rules ask for:
+`@levels` is a `D_tree_api`-style second copy of a list slot, and hook-owned
+state is *synced from an invariant*, not toggled by the hooks (`D_progress_bar`'s
+`sync_ticker`). Per-level truncate closures wired at `push` are the toggle
+version.
+
+**Left button only.** `MouseEvent` is X10 press-only (no release, no motion), so
+there is no drag case. Excluding scroll is `D_notification`'s stray-spin lesson;
+excluding `:right` keeps a future context action from nuking an open dropdown.
+
+**Vaadin, verified against the 24 docs.** "Modal dialogs are closable in three
+ways: by pressing Esc; clicking outside the Dialog; or programmatically", and
+"Dialogs are modal by default" — so default-`true`-even-for-modals is the Vaadin
+behavior, and that is why the default is what it is. What does *not* port: in
+Vaadin the thing catching the outside click is the modality curtain, part of the
+overlay, so light dismiss is nearly free because modality is a DOM element.
+Tuile's modality is a routing rule with nothing to click on, so the notice must
+be manufactured. Which also means Vaadin's *non-modal* behavior is no precedent
+here — a non-modal Vaadin Dialog does not light-dismiss; its ComboBox overlay
+does.
+
+**The modal/non-modal split dissolves.** The design was framed as two halves,
+only one with a customer: non-modal overlays needing a notice, and modals unable
+to hear a click at all. The flag applies identically to both, and no
+`clicked ||= modal_popup` routing change is needed, because nothing is
+*delivered* to the modal — it is just closed. An outside click on a modal both
+dismisses it and is swallowed (click once to dismiss, again to act), same as
+Vaadin's curtain.
+
+**Roads not taken.** A veto — `on_close` (or a new hook) returning false to
+refuse the close: rejected mechanically, since `on_close` fires from
+`on_detached`, after the popup is off the screen, and you cannot un-detach. Any
+veto therefore needs a *new*, earlier hook, which is the notice again with a
+return channel, and it makes every grouped overlay re-implement the geometry
+test the pane just did. A `Screen`-level "a click landed at P" broadcast any
+component can subscribe to: rejected on sight, a second mouse-dispatch path
+beside the one-chain rule. Making `ListDropdown` modal so it hears every click:
+`ComboBox` and `Select` would lose the events their own faces need. A generation
+counter to make close-and-reopen-within-one-click safe: over-engineering for a
+case nothing hits — reopening the *same* popup object during delivery of one
+click is out of contract (the snapshot holds it), and the answer is a fresh popup
+or a cleared flag.
+
+**Per-widget settings.** `Popup` defaults `true`; `ListDropdown` inherits it, so
+`Select`, `ComboBox` and every cascade panel are fixed with zero wiring;
+`Notification` sets `false`, since a toast is timed and an unrelated click is not
+about it; app modals keep `true` and opt out per dialog. The one accepted risk is
+a stray click discarding a half-filled form dialog.
+
+## D_no_context_menu — No `ContextMenu`: designed, priced and declined (2026-08-24)
+
+**Status:** Decided 2026-08-24 — **not building it**, indefinitely. Designed in a
+since-retired `ideas/context-menu.md` (opened and graduated the same day), so
+this entry is the whole record. `Context Menu` was *dropped* from
+`ideas/new-components.md` rather than demoted to its Tier 3, and nothing else
+tracks it.
+
+**Context.** The roadmap listed it as a Tier 1 near-freebie — "same as Menu Bar;
+`:right` already parses" — and after `MenuBar` shipped that looked right: the item
+tree, the mnemonics, the cascading submenus and the per-level width measurement
+all exist and would have been reused as they stand. The design confirmed it. The
+widget then failed on its *inputs*, not on its machinery, which is why this entry
+is a rejection rather than a deferral.
+
+**Decision, and the three reasons in order of weight.**
+
+1. **The gesture that defines the widget is the least reliable input Tuile has.**
+   A context menu *is* right-click, and terminal emulators routinely keep that
+   button for their own menu (some pass it through only with Shift) — on top of
+   mouse reporting being optional in the first place. The keyboard route then has
+   to be invented from nothing: no terminal sends a context-menu event, where a
+   browser hands Vaadin `contextmenu` from Shift+F10 *and* the Menu key, so
+   Vaadin's `ContextMenu` needs no keyboard code at all. **And Shift+F10 is not
+   readable today:** `Keys.getkey` gulps at most 5 bytes after `\e` — deliberately,
+   since 6 would over-read the next event on a mouse burst — while xterm sends
+   `\e[21;2~`, 6 tail bytes, so the `~` would surface as a printable keypress.
+   `\e[29~` (Menu/Apps) and plain F1–F12 *do* fit. That constraint binds anything
+   wanting an exotic key, not just menus.
+2. **No host wants one.** Not the sampler, not `file_commander`, and the TUI
+   lineages are thin: mc spends F9 on a menu bar instead, Turbo Vision and LazyGit
+   have none. LazyVim is the counterexample — it does ship one — which is an
+   argument for revisiting when a host asks, not for building on spec.
+3. **It would cost two new framework concepts to serve nobody** — an invisible
+   modal popup as a focus grab, and a `ScreenPane` notice for modality-blocked
+   clicks (the second outlived it; see below).
+
+**The design that would have been built,** recorded so a revival starts here. One
+structural fact drives all of it: **a popup can only hold focus if it is modal.**
+`ScreenPane#handle_key` scopes delivery to `modal_popup || content`, so a *focused
+non-modal* popup sits outside the key scope and every keystroke goes dead —
+AGENTS.md's non-modal-overlay trap. There is no third option, and unlike a menu
+bar a context menu has no strip to park focus on.
+
+So: `ContextMenu < Popup(modal: true)` with a **zero-size rect that paints
+nothing** — not a picture but a *grab*, playing exactly the role `MenuBar`'s strip
+plays (focus holder, key scope, lifecycle owner, outside-click sink). Every
+visible panel, level 0 included, is a `Cascade` level, so `Cascade` and `Item` are
+reused verbatim and mnemonics work with no new code at all. Modality then hands
+over focus save/restore (`@popup_prior_focus`), an inert Tab (`cycle_focus` scopes
+stops to `modal_popup`, and a grab has none) and click-blocking for free. Two
+openers, because the desktop lineages agree these are different placements:
+`open_at(point)` for the mouse, `open_below(rect)` for the keyboard — pointer
+versus selection. Framework growth: `ListDropdown#anchor_at(point)`, a second
+level-0 entry point on `Cascade`, the blocked-click notice, and overrides for
+`reposition` (or close-on-resize, as `MenuBar#rect=` does), for `q`/ESC — `q` has
+to stay available as a mnemonic — and for `keyboard_hint`.
+
+**Alternatives rejected.**
+
+- **Host-driven, no new machinery** — a plain object the host wires from its own
+  `handle_mouse` / `handle_key`, i.e. `MenuBar`'s architecture minus the
+  component. It costs the framework nothing, and that is the trap: `MenuBar`
+  encodes five invariants *once* because it is a component — close on focus loss,
+  on detach, on resize, swallow keys while open, forward the mouse — and every
+  host would re-encode all five. Forgetting `on_detached` strands panels on the
+  pane with nothing to take them down, the exact bug class AGENTS.md's
+  non-modal-overlay section exists to prevent.
+- **The level-0 panel *as* the modal popup**, which deletes the invisible
+  component. Rejected because level 0 then becomes structurally unlike every
+  deeper level, so the panel-driving logic — `MOVE_KEYS` to the highlight, Enter
+  to drill-or-fire, mnemonic match, truncate-on-cursor-move — exists twice for
+  panels that are identical on screen. It buys only the deletion of a zero-size
+  rect.
+- **Recursive modal popups, one per level, no `Cascade`** — each level an ordinary
+  modal `Popup` over a *focusable* `List`, with `Popup`'s own ESC/`q` closing a
+  level. Genuinely tiny and free of every non-modal trap, and rejected on the
+  smell: it is a *second* menu mechanism, so item trees, mnemonics, submenu
+  arrows, width measurement and the key map would all get a second
+  implementation. If it is right, `MenuBar` is wrong — a much larger argument
+  than this widget.
+- **A `Component#context_menu=` slot** checked inside `Component#handle_key`, so
+  any component gets one by assignment. Half a feature: almost no widget calls
+  `super` from its own `handle_key` (`List` doesn't), so it would work for
+  ancestors that don't override and silently not for focused leaves.
+- **Vaadin's `setTarget(component)`** — attach the menu to a target and let the
+  framework route the right-click to it. Tuile has nothing to build that on:
+  `handle_mouse` returns `void`, and a right-click already reaches *every*
+  component along the rect chain, ancestor first and deepest last, so "which
+  target owns this click" has no answer. (What that ordering *would* give free is
+  deepest-wins, if a revival adds "opening one closes any other open context
+  menu" — the `D_notification` shape, found by scanning the popups stack rather
+  than a class ivar.)
+- **Type-ahead search inside an open menu**, which `List#select_next` makes nearly
+  free. Same rejection as in `D_menu_bar`: it competes with explicit mnemonics for
+  the same keystroke and owes a precedence rule.
+
+**Two gaps it surfaced that outlive it.**
+
+- **An outside click on an open overlay notified nobody.** `Select`, `MenuBar`
+  and the sampler's slash menu all lingered on a click that landed on decoration,
+  and a modal popup could not dismiss on an outside click at all. **Closed**
+  2026-08-24 by `D_outside_click`, which also dissolved the modal/non-modal split
+  the gap was framed around.
+- **A right-click does not move a `List` cursor.** `List::Cursor#handle_mouse`
+  acts on `:left` only (specced), and there is no public `item_index_at(point)`,
+  so "act on the row I clicked" is unsayable unless the app computes
+  `event.y - rect.top + scroll_top_row` itself. Nothing needs it today; it is the
+  same shape of hole as the `List#select(index)` gap `D_menu_bar` had to fill.
+
+---
+
+## D_status_bar — Delete the framework status bar; the app owns its bottom row (2026-08-25)
+
+**Status:** Accepted 2026-08-25; unimplemented. Supersedes the shipped
+`ScreenPane#status_bar` slot and the `Component#keyboard_hint` channel that fed
+it — see *the scar* at the end. Retires `ideas/status-bar-ownership.md`.
+
+**Context.** `ScreenPane` has always reserved the bottom terminal row for a
+framework-owned `Label`, and `Screen#refresh_status_bar` filled it on every
+focus change from three sources: a hardcoded `"q quit"`, the `hint:` strings on
+registered global shortcuts, and one component's `keyboard_hint` — the innermost
+active `Window` (found by an `is_a?` scan) when tiled, the top popup's *direct*
+content when not.
+
+That last source barely worked. Of the seven `keyboard_hint` implementations,
+only `Window`, `Popup` and `PickerWindow` (a `Window`) were reachable in any
+configuration; `MenuBar`, `Tabs`, `Select` and `ComboBox` were dead
+**everywhere**, tiled and popup alike, because nothing walked down to the
+focused component and the popup path forwarded only to its direct child. The
+obvious fix — ask `screen.focused` and walk up, matching the delivery bubble —
+was drafted, and a survey of the two real consumers was run to choose between
+it and two variants. The survey concluded the channel should be **deleted**.
+
+**Decision.** Delete the status bar and the hint channel. `ScreenPane` no longer
+owns a `Label`, no longer reserves `height - 1`, and `Component#keyboard_hint`
+ceases to exist. In its place `Screen` gains one notification —
+`on_focus_changed=`, a plain proc fired from `focused=`, matching the
+`on_theme_changed=` style stock assemblies already use. An app that wants a
+status bar builds one:
+
+```ruby
+bar = Tuile::Component::Label.new
+root = Tuile::Component::Layout::Vertical.new
+root.add(main, Expand)
+root.add(bar, Fixed[1])
+screen.on_focus_changed = -> { bar.text = hint_for(screen.focused) }
+```
+
+**Why deletion beat a better hint source.**
+
+- **No app has ever wanted a *widget's* hint.** Across four apps, virtui
+  advertises window-level app keys (`"p Power  v run Viewer  m Memory  d toggle
+  Disk stat  / Search"`) and pikuri-tui advertises global app keys (`"^K menu"`,
+  `"^C cancel"`). Neither has ever advertised a `Select`'s or `ComboBox`'s keys.
+  The channel was not merely unused by four of its seven implementors — the
+  thing it was designed to carry is something nobody wants carried.
+- **An app was routing presentation through dispatch.** pikuri re-registers a
+  global keybinding to change a status-bar string, and documents the technique
+  in rdoc: "`Screen` replaces the binding in place on re-register, so this is
+  also how the hint stays in sync with the counter." The bar was write-only from
+  the app's side, so a *text* change had to be expressed as a *binding* change.
+  That is the design inverted, not a missing feature — and it is the single
+  finding that settled this.
+- **The one reachable widget hint was also stale.** `MenuBar#keyboard_hint`
+  switched to `"↑↓ move  ⏎ select"` with the cascade open, but the cascade is a
+  non-focusable `ListDropdown`, so focus never changed and `refresh_status_bar`
+  never ran (it fired from `focused=`, `theme=` and the two registry mutators —
+  never from `add_popup`). Opening a menu did not update the bar; *closing* it
+  did, via focus repair. Dead twice over.
+- **The bar is a layout special case that `Box` layouts obsoleted.** The
+  `height - 1` reservation is v0.1-era, from before `Vertical`/`Fixed` existed.
+  An app-owned bar is now three lines, and buys what the framework can never
+  offer: two rows, a bar at the top, its own styling, a file-commander
+  function-key strip, or nothing at all.
+- **It is the shape the top-down re-grow rule already governs.** That rule says
+  a deleted bottom-up channel may return only as an *optional, read-only,
+  caller-side query*, never as an automatic channel the framework consults.
+  `keyboard_hint` was an automatic channel; deleting it applies the rule Tuile
+  already lives by.
+- **The framework baked an app policy.** The `"q quit"` prefix was
+  unconditional: pikuri's three apps quit via `^K → q`, and their bar read
+  `q quit  ^K menu` while `q` typed into the focused input just typed a `q`.
+
+**Alternatives rejected.**
+
+- *Walk the focus chain and concatenate (the drafted fix).* Correct as far as it
+  went — it matched the delivery bubble, subsumed the popup special case, and
+  would have deleted `active_window`. Rejected because it fixes *reachability*
+  while leaving ownership where it hurts: pikuri's re-registration hack survives
+  it untouched, and MenuBar's flickering, redundant `←→ menu  ⏎ open` becomes
+  *visible* rather than merely dead. It also forced a ruling on hint ordering
+  that is really a truncation policy, since `Label` ellipsizes and the rightmost
+  hint silently vanishes on a narrow terminal.
+- *Ask `active_window` and forward down the active chain.* Keeps `Window` as the
+  unit of "what am I looking at" but re-implements the focus walk, and preserves
+  the framework's only place where a *class* is special-cased for behavior.
+- *Keep the bar, make it optional.* A `status_bar: false` flag leaves every
+  defect in place for whoever leaves it on, and adds framework surface in the
+  middle of an argument for less of it.
+- *Drop only `MenuBar#keyboard_hint`.* Treats the symptom. Three other widget
+  hints stay dead, and the ownership inversion is untouched.
+- *Keep `Component#keyboard_hint` as a documented seam, delete only the
+  renderer.* Tempting — it preserves a common vocabulary for a future component
+  ecosystem. Rejected for now because a seam with no framework consumer is
+  precisely the automatic-channel-with-no-caller the re-grow rule exists to
+  prevent, and because the built-in hints it would preserve are the four nobody
+  wants. See the re-grow shape below.
+
+**Consequences — what was given up, honestly.**
+
+- **Zero-config batteries are gone.** `book/01-first-app.md` said "you never
+  created a status bar, yet the app has one", and `hello_world_spec` asserted on
+  `q quit`. A first app now shows an empty bottom row until it builds one. Ruled
+  acceptable: a bar the app cannot drive is not a battery, and ch1 gains a
+  better story once the bar is three lines of `Vertical`.
+- **A widget's keys are no longer self-describing.** An app that *does* want to
+  advertise a `ComboBox`'s keys must hardcode `"↑↓ select  ⏎ accept"` itself,
+  duplicating knowledge that lived in the widget. No app has ever done this, but
+  the duplication is real if one starts.
+- **`book/05-focus.md`'s "The status bar writes itself" section goes.** It
+  claimed the bar was "driven by focus" and showed "the focused context's own
+  advertised hint" — behavior that never existed; focus only triggered the
+  rebuild. Deleting it removes a documented promise the code never kept.
+- **`D_boolean_fields`' aside is retired**, not overruled: "hints are a
+  window/popup-level affordance; per-field hints would drown the status bar" was
+  an argument about where a hint belongs, and there is no longer a framework
+  hint to place.
+- **A modal {Component::Popup} no longer shows how to close itself.**
+  `Popup#keyboard_hint`'s `q Close` was the only affordance, and — unlike
+  `PickerWindow`'s hint, which merely repeated the option keys its own `List`
+  rows already paint — nothing else on screen carries it. `popup.rb`'s `q`/ESC
+  handler is untouched, so the behavior remains; only the advertisement is gone.
+  **Ruled acceptable 2026-08-25 on the Vaadin precedent:** a Vaadin `Dialog`
+  closes on ESC and no Vaadin *app* documents that anywhere — it lives in the
+  framework's own docs and javadoc, which end users never read. ESC-dismisses-an
+  -overlay is a convention the user brings with them, not something each app has
+  to teach. An app that wants it spelled out writes it into its own row.
+
+**The `q`/ESC quit fallback stays** (`Screen#event_loop`:
+`@event_queue.stop if !handled && ["q", Keys::ESC].include?(key)`). It is the
+same baked app policy as the `"q quit"` string, but it is *dispatch*, not
+presentation, and it is separable — deleting it would make every example and
+both downstream apps grow a quit handler in the same breath as an unrelated
+change. **Ruled 2026-08-25 by `D_quit_key`: it stays, unadvertised**, on the
+same convention argument as the popup's lost `q Close` above.
+
+**There is no app-facing `keyboard_hint` convention, and the book must not
+teach one.** The first cut of `examples/file_commander.rb` kept a
+`PaneWindow#keyboard_hint` and walked up the focus chain via
+`respond_to?(:keyboard_hint)` to find it — which re-created the deleted seam by
+convention, in three places at once (the example, the book, virtui), with a
+duck-type where a declared method used to be. It was also *dead*: both panes
+were `PaneWindow`s returning the same constant, so the focus hook, the walk and
+the duck-type together computed a value that never changed. The example is now
+a static `Label` and `PaneWindow` is gone; book ch5 leads with "a status line is
+a `Label` in your layout", and treats {Screen#on_focus_changed=} as the
+*exception* for a row that genuinely varies. The walk survives only in virtui,
+where three windows really do advertise different keys — as one app's design
+decision, named as such.
+
+**Re-grow rule.** A hint channel may come back only as **a query the app pulls,
+never a channel the framework pushes** — and specifically not as a
+framework-owned row. Textual is the shape to copy if it does: its `Footer` is a
+widget the app mounts in `compose()`, reading from the `BINDINGS` table the
+framework owns ⚠. That splits ownership at the right seam — the app decides
+whether a bar exists and where, the widget declares its keys — and it is already
+on record as steal-candidate #1 in `D_key_dispatch`. Bringing back a bar the
+framework *places* reopens this entry.
+
+**Prior art** (surveyed 2026-08-25; ⚠ marks memory-based claims worth checking
+before acting). The honest reading is that a framework-owned status *row* is a
+minority position, and the one framework that does it well does not own the row:
+
+| | Owns a status row? | Where the text comes from |
+|---|---|---|
+| **Turbo Vision** | yes — `TStatusLine`, always present | declarative `TStatusDef` tables keyed by help context ⚠ |
+| **Textual** | no — `Footer` is a widget you mount | the framework's `BINDINGS` tables ⚠ |
+| **Swing** | no | app-written `JLabel` in `BorderLayout.SOUTH` |
+| **ncurses / Bubbletea / Ratatui** | no | app draws every cell |
+| **Tuile (before)** | yes — `ScreenPane#status_bar` | `active_window&.keyboard_hint` + registry hints + `"q quit"` |
+| **Tuile (after)** | no | app-drawn, from `on_focus_changed` |
+
+Turbo Vision is the only real precedent for the shipped design, and it paired
+the row with a declarative binding table — the half Tuile never had, which is
+why its bar could only be fed by an inverted registration hack.
+
+**The scar.** The status bar was never designed for Tuile. It arrived whole in
+`4491a77`, the 0.1.0 commit that ported virtui's `lib/ttyui/` under the `Tuile`
+namespace — it was *virtui's* status bar, generalized by accident of extraction,
+and virtui is to this day the only app using the `keyboard_hint` half. It then
+survived every later overhaul (the top-down layout rewrite, the key-ladder
+deletion, the tree-first split) without anyone asking who it was for, while each
+new widget dutifully grew a hint nobody could see. The tell sat in the code the
+whole time: `Screen#active_window` was public API with exactly one caller —
+this one — and no app ever invoked it.
+
+---
+
+## D_quit_key — `q` / ESC quit the loop, unadvertised, as a Tuile quirk (2026-08-25)
+
+**Status:** Accepted 2026-08-25; no code change — this records a decision to
+*keep* what ships. Closes the question `D_status_bar` deferred.
+
+**Context.** `Screen#event_loop` ends with
+`@event_queue.stop if !handled && ["q", Keys::ESC].include?(key)` — after the
+three-rung ladder has declined a key, bare `q` or ESC stops the loop and the
+app exits. It is app policy the framework enforces, and no app opted into it.
+
+`D_status_bar` deleted the framework status bar and with it the hardcoded
+`"q quit"` prefix that was this fallback's only advertisement, deliberately
+leaving the behavior alone as a separate question. That left the least coherent
+state of the three: a hardcoded quit key with nothing anywhere surfacing it.
+
+**Decision.** Keep it exactly as it is, unadvertised, and stop treating it as an
+open question.
+
+- **It is a convention, not an invention.** `q` quits `less`, `man`, `top`,
+  `htop` and every pager git shells out to; ESC dismisses. A user arriving at a
+  full-screen terminal app already tries both. That is the same argument that
+  settled the popup's lost `q Close` hint in `D_status_bar` — a convention the
+  user brings is not something each app must teach.
+- **The escape hatch already exists and needs no new surface.** A component
+  keeps `q` by consuming it, which is the whole of `D_key_dispatch`'s
+  delivery rung: a focused {Component::TextField} does it for free (`q` is
+  printable — this is why pikuri-tui's shells never quit on a typed `q`), and an
+  app wanting `q` as a command binds it in the scope root's `handle_key`. ESC
+  likewise never reaches the loop while a {Component::Popup} is open, because
+  the popup consumes it first.
+- **It is genuinely useful for the small app.** `examples/hello_world.rb` is
+  eleven lines and needs no quit handler. Deleting the fallback would make every
+  example and both downstream apps grow one, buying nothing.
+
+**Alternatives rejected.**
+
+- *Delete it; apps handle their own quit.* The clean-architecture answer, and
+  the one consistent with deleting the status bar. Rejected because the two are
+  not the same shape: the status bar was a *row the app could not write to* —
+  it actively blocked apps (pikuri had to re-register a keybinding to change
+  text) — whereas this fallback blocks nothing. Any component can take the key.
+  A rule the app can override on the spot is a default, not a policy.
+- *Make it opt-in (`Screen#quit_on_q=`).* Adds framework surface for a knob
+  nobody has asked for, in the middle of an argument for less of it, and the
+  override it provides is one the key ladder already gives for free.
+- *Re-advertise it somehow.* That is the framework-owned status row again.
+
+**Consequences.**
+
+- **It is undiscoverable from inside the app**, and that is accepted. An app
+  that wants it spelled out writes `q quit` into its own status line —
+  `examples/hello_world.rb`, `examples/file_commander.rb` and virtui all do;
+  pikuri-tui deliberately does not, because its focused input eats `q` and the
+  hint would be a lie.
+- **`q` is reserved-ish for a scope root.** An app binding bare `q` in
+  `handle_key` must return `true`, or the key falls through and quits the app —
+  a surprising bug the book calls out (ch5) and this entry pins.
+- **What would reopen it:** a real app that needs bare `q` at the scope root and
+  finds consuming it awkward, or a second key wanting the same treatment (which
+  would make this a *list*, and a list wants a knob).
+
+---
+
+## D_hook_visibility — A framework-invoked hook is protected, reached with `__send__` (2026-08-30)
+
+**Status:** Accepted and implemented for {Component#on_theme_changed}
+(protected; `Screen#theme=` fans it out through `__send__`). `Component#on_focus`
+is the one framework-invoked hook still public — see *Consequences*.
+
+**Context — the field report.** virtui crashed on an OS appearance flip:
+
+```
+NoMethodError: protected method `on_theme_changed' called for an instance of UI::VMWindow
+```
+
+Three of its `Window` subclasses group their overrides together —
+`on_width_changed`, `on_theme_changed`, `repaint_border` — under one `protected`
+keyword. Two of those three are protected in Tuile; the third was **public**,
+because `Screen#theme=` fanned it out as `@pane&.on_tree(&:on_theme_changed)`, an
+explicit-receiver send. Ruby lets a subclass *narrow* an inherited method, so the
+natural grouping silently broke the walk.
+
+The failure is worse than one exception. `theme=` assigns `@theme` *before* the
+walk, so the theme really does swap; the walk then dies at the first offender in
+pre-order, every component after it never hears the hook, and the closing
+`needs_full_repaint` never runs — the new theme is live under content painted
+for the old one, until something unrelated invalidates. And nothing catches it in
+a test suite that never flips the theme.
+
+**Decision.** A hook the *framework* calls on a component is plumbing an app
+overrides and never invokes, so it is `protected`, and the framework reaches it
+with `__send__`:
+
+```ruby
+@pane&.on_tree { _1.__send__(:on_theme_changed) }
+```
+
+`__send__` is the point, not a workaround for the visibility change: it ignores
+visibility, so an override may be public, protected or private and the walk can
+no longer be broken from an app at all. The public half of the seam is the
+*listener* (`on_theme_changed=`), which is what an app assembling stock
+components actually calls.
+
+**Why the visibility can't just be finessed by *who* does the walking.** Ruby
+checks a protected call against the class the method is *defined in*, relative to
+the caller's `self` — so an override defined in a subclass is unreachable by
+explicit receiver from anywhere else, including a sibling component and even the
+base class:
+
+```ruby
+class Base;  def fan(o) = o.hook; protected; def hook = "base"; end
+class Sub  < Base; protected; def hook = "sub";  end
+class Other < Base; end
+Other.new.fan(Sub.new)   # NoMethodError
+Base.new.fan(Sub.new)    # NoMethodError — same reason
+```
+
+That leaves exactly two workable shapes: `__send__`, or an *implicit* receiver.
+{Component#fire_lifecycle} is the implicit-receiver one — it recurses through a
+`Component`-defined method and calls `on_attached` / `on_detached` on `self`,
+which is why those two have been quietly protected all along and why this class
+of bug never reached them.
+
+**Alternatives rejected.**
+
+- *Keep the hook public and document "don't narrow it".* The documentation
+  nobody reads, guarding a trap the natural code layout walks straight into —
+  virtui's three sites are the proof, and grouping hooks under one `protected` is
+  good Ruby, not a mistake to correct. A rule that fires an exception in
+  production, at OS-flip time, in a path no suite exercises, is not a rule; it is
+  a landmine.
+- *A public `Component#fire_theme_changed` walker* — the `fire_lifecycle` shape,
+  hoisted to public so `Screen` can start it, calling the protected hook on
+  `self` at each node. It genuinely works and needs no `__send__`. Rejected on
+  surface: it puts a second public method on every component (in the rdoc, in
+  `sig/tuile.rbs`, callable by apps) and duplicates {Component#on_tree}, to avoid
+  one `__send__` at one call site. It also only relocates the hazard — the
+  *walker* becomes the method that must not be narrowed.
+- *Rescue `NoMethodError` around the walk.* Swallows real bugs inside app hooks
+  and leaves the restyle half-applied, which is the symptom being fixed.
+- *Make the call tolerant but leave the hook public.* Fixes today's crash and
+  keeps the trap armed for the next contributor who writes `&:some_hook`; the
+  hook's visibility is what states the intent.
+
+**Consequences.**
+
+- **The `attr_writer` stays public**, and the reader is now protected — an
+  asymmetric accessor pair, deliberately: assigning a listener is app-facing,
+  firing it is not.
+- **A new framework-invoked hook copies this shape**: protected, `__send__` at
+  the fan-out, listener writer public if it has one. Never `&:hook`.
+- **`on_focus` stays public, deliberately** — it is not plumbing in the same
+  sense. {Component::HasContent} / {Component::Layout} / {Component::TabSheet}
+  each override it to forward focus into their content, so it reads as part of
+  the composition seam a mixin publishes rather than as a private notification.
+  It carries the same theoretical hazard (an app narrowing its override would
+  break `Screen#focused=`), and that is accepted: nothing has hit it, and
+  protecting a method three mixins present as interface would cost more clarity
+  than it buys.
+- **Specs call the hook with `send`**, and two guards exist: `component_spec`
+  asserts the visibility pair, `screen_spec` asserts that a subclass declaring a
+  `protected` override is still fired *and* that the walk continues past it.
+
+## D_overlay — Extract `Overlay`; `Popup` becomes unconditionally modal (2026-08-30)
+
+**Decision.** Split {Component::Popup} in two. `Overlay < Component` is the bare
+floating layer — the mount/dismiss lifecycle, `owner`, `on_close`,
+`close_on_outside_click`, a no-op `reposition`, and the full-repaint escalation
+in `rect=` — and `Popup < Overlay` adds the modal dialog on top: a declared
+`size`, self-centering, `focusable?`, and ESC/`q`. `Popup.new(modal: false)` is
+gone; the `@modal` ivar with it, since `modal?` is now a constant on each class.
+`Notification` and `ListDropdown` both reparent onto `Overlay`.
+
+**Why.** The cut line was not invented — it is exactly what `ScreenPane` calls on
+a member of `@popups` (`rect`, `modal?`, `reposition`, `owner`,
+`close_on_outside_click?`, `close`, `on_tree`), so `Overlay` makes an interface
+that already existed implicitly into a class. What forced it was the tally: both
+non-modal subclasses *rejected* most of `Popup`. `Notification` overrode
+`focusable?`, `tab_stop?`, `reposition` and `handle_mouse`, and had to **raise**
+from `size=` to fight off an inherited feature; `ListDropdown` called `self.size
+=` only to stop `Popup#reposition` stomping its anchored placement. A base whose
+contract is "remember to switch four inherited behaviours off" fails the `cop`
+skill's *the base must earn its place — it permits, it doesn't mandate*, and this
+file's own AGENTS.md section "Non-modal overlays — two traps a new one will hit"
+was that fragile-base-class tax written down in prose because it could not be
+written in types.
+
+**What it bought, beyond tidiness.**
+- *A latent bug, fixed by construction.* `ListDropdown` declared
+  `focusable?`/`tab_stop?` on its inner `Menu` and never on itself, so it
+  inherited `Popup#focusable? == true`. It survived only because `layout` makes
+  the content cover the whole rect, so `HasContent#handle_mouse` forwarded every
+  in-rect click before `Component#handle_mouse` could assign focus — safe by
+  *geometry*, not by declaration. A border or an inset would have landed focus
+  outside the key scope and killed every keystroke until Tab.
+- *`Notification`'s raising `size=` is deleted, not renamed.* An `Overlay` has no
+  declared box, so there is nothing to refuse.
+- *`anchor_to` stops double-assigning `rect`.* The `self.size =` call ran through
+  `Popup#reposition`, which assigned an intermediate rect at the old origin —
+  reintroducing exactly what `reposition`'s own rdoc says it avoids, on
+  `ComboBox`'s per-keystroke re-anchor path.
+
+**Alternatives rejected.**
+- *The inverse cut — base stays `Popup`, the modal one becomes `Dialog`.* It
+  would have kept `ScreenPane#popups` accurate at zero renaming cost, but every
+  existing `Popup.new(content: w).open` would **silently** become non-modal,
+  uncentered and un-ESC-able. A loud `ArgumentError` on a removed `modal: false`
+  beats a silent behaviour change.
+- *Renaming the pane/screen vocabulary to `overlays` / `add_overlay`.* Considered
+  and declined: the break reaches `Screen#add_popup`, which apps call, and
+  "popup stack" remains a defensible name for the stack. The `@param` types and
+  rdoc say `Overlay` instead, which is what a caller actually needs to know.
+- *Keeping `modal:` as a kwarg on `Popup` alongside `Overlay`.* Two ways to build
+  the same thing, with the trap-laden one still reachable. The whole point is
+  that the inert defaults are the ones you get by default.
+
+**Consequences.**
+- `focusable?` and `modal?` are now coupled: flip both or neither. A *focusable
+  non-modal* overlay is the one combination that must never ship — it holds focus
+  outside the key scope, where `bubble_key` reaches nobody. The `Overlay` rdoc
+  states this as the coupling rather than as a ban on overriding, because `Popup`
+  overrides both.
+- `Overlay#reposition` is a no-op, so a subclass with a *derived* position owns
+  its own override (`Notification`), and one placed by a driver simply keeps the
+  rect it was given (`ListDropdown`, re-anchored from its driver's `rect=`).
+- **Still open:** where `anchor_to` / `anchor_beside` belong. They stay on
+  `ListDropdown` for now. The `Popover` extraction (`D_select`, `D_menu_bar`) is
+  *cheaper* after this change, since `Overlay` — not the modality-carrying
+  `Popup` — is the right parent for a generically anchored layer; the trigger is
+  unchanged, the first non-`List` content wanting anchoring.
+
+## D_declared_size — `Popup#size` is `#declared_size` (2026-08-30)
+
+**Decision.** Rename `Popup#size` / `#size=` — and the `Popup.new` and
+`InfoWindow.open` keyword — to `declared_size`. Nothing else changes: it is still
+a `Size | Fraction`, still re-resolved against the screen on every layout pass,
+still authoritative.
+
+**Why.** `size` on a component is reasonably expected to mean `rect.size`, and
+the name was already spoken for by a different concept: an *input* the popup is
+re-read from on every layout pass, not a *report* of current geometry. Leaving it
+would have created tension the first time anyone added `Component#size`, and the
+types disagree too — `rect.size` is a `Size`, this is a `Size | Fraction`, so a
+`Component#size` reader would have been a Liskov break on the one class most
+likely to be handled polymorphically. `declared_size` is the word `popup.rb`'s
+own rdoc already used ("its box is *declared* by"), and it contrasts correctly
+with `preferred` / `requested`, which the same rdoc explicitly disclaims: the
+screen applies exactly what you ask for, with no negotiation.
+
+**Alternatives rejected.**
+- *`auto_size`* (the first proposal). Two problems. "Auto-size" conventionally
+  means shrink-to-fit-content — Swing's `pack()`, WPF's `SizeToContent` — which
+  is precisely the eager bottom-up `content_size` channel deleted in 0.9.0, and
+  which AGENTS.md already spends a rule keeping out of the vocabulary
+  (`D_box_layouts`: "there is no `Auto`"). And it read as a contradiction on the
+  one subclass that genuinely does size itself: `Notification#auto_size=` raising
+  "sizes itself from its messages". (That override is gone under `D_overlay`, but
+  the naming argument stands for the next such subclass.)
+- *`auto_center_with_size(x)`*, a command rather than a property. It has real
+  merit — Tuile already has imperative geometry methods (`center`, `reposition`,
+  `anchor_to`) and a command is honest about the side effect that a bare setter
+  hides. Rejected on three counts: it collides with the existing `Popup#center`,
+  giving two near-synonymous centering verbs; after `D_overlay` made `Popup`
+  unconditionally modal, "auto center" names the class *invariant* rather than
+  the varying member, so it carries no information at the call site; and the
+  member is *state*, not an action — `reposition` re-reads it forever, so a
+  `Fraction` means "stay half the screen through every SIGWINCH", which a command
+  name hides. A rejected refinement, folding it into `center(size = nil)`, keeps
+  one verb but still hides the persistence.
+- *Keeping `size` and never adding `Component#size`* — settling that a
+  component's size is spelled `component.rect.size` forever. Cheapest (no
+  breaking change), and declined because it preserves the trap rather than
+  removing it.
+
+**Consequence, taken up the same day.** `size` being free on `Component` was the
+point, and `Component#size` / `#width` / `#height` were added straight after as
+pure readers of `rect` — partly to *squat* the names, so no component can later
+claim `size` for a content-derived measurement. They are reports, never requests:
+no writer, and no container consults them when dividing space. The top-down
+re-grow rule still governs, and a second component wanting a *declared* box
+copies `Popup`'s naming rather than overloading `size`.
+
+## D_extent — `extent` on `Component`: the rect is what you were given, not what you paint (2026-08-30)
+
+**Decision.** Promote `extent` from a per-widget convention to a `Component`
+member defaulting to `rect`, and give it a paired `clear_outside_extent`. A
+widget that paints less than its rect narrows the extent and then uses it in the
+three places that care — clearing, hit-testing, anchoring. **`rect` keeps meaning
+exactly what the parent assigned**; nothing about `extent` flows upward.
+
+**Why now.** The concept had leaked six times (`Button`, `Checkbox`, `Tabs`,
+`MenuBar`, `Select`, `ComboBox`) and produced two shipped bugs in one session: a
+`Select` in a single-slot container opened its dropdown from a click 20 rows
+below its face, and `ListDropdown#anchor_to` placed the panel using a rect the
+driver did not occupy. Both were "the widget forgot to consult `extent`", in
+different places. It was one concept with no name and no home.
+
+**Rejected: the component clamps its own `rect`.** The tempting inverse — layout
+offers 10 rows, the widget writes back 1, the parent's arithmetic unaffected and
+the gap simply showing through. It is *not* the deleted bottom-up `content_size`
+channel (no parent consults anything, no re-layout is triggered), and
+`children_tile_rect?` already handles the resulting gap, so it would have worked.
+Rejected on the invariant it costs: **`rect=` would lie.** `c.rect = r; c.rect ==
+r` becomes false, and `f.add(checkbox, Fixed[10])` silently yields a one-row
+checkbox, so the declared *constraint* is a lie too, invisible at the call site.
+In a top-down system the highest-value property is that a container's arithmetic
+can be verified by reading the container alone; once any component may rewrite
+its rect, you cannot reason about a `Box` without knowing which of its children
+clamp. Two lesser counts: clamping spreads the surprise to every reader of any
+`rect` (with `extent` only the six widgets that *have* a quirk carry it), and it
+solves only the height axis — every width clamp here is content-derived
+(`caption.display_width + 4`), so it would need the parent told to re-lay-out on
+`caption=`, which is `on_child_content_size_changed`, deleted in 0.9.0. Vaadin 8's
+slot negotiation is the prior art, and Vaadin 10 dropped it for CSS.
+
+**The default is `nil`, not `rect.size`, and that is what lets `repaint` decide.**
+The first cut defaulted to `rect.size` and had the base branch on `extent ==
+rect`; it was implemented and backed out, because a one-row `Select` in a one-row
+rect satisfies that test while genuinely painting its extent in full — the base
+would blank the row and `Select` would repaint it, every cell dirty, the row
+re-emitted (`D_progress_bar`). The two states the base must tell apart are "no
+declaration, so clear everything" (a `Label` with short text) and "declared, so
+leave it alone", and they are *not* distinguishable from the value: they are
+distinguishable by whether there is a value. Hence `nil`.
+
+The payoff is that widgets keep the ordinary `super`-then-paint shape — no widget
+has to remember to call a helper, and forgetting to declare an extent degrades to
+today's behaviour rather than to stale glyphs.
+
+**`extent` is a `Size | nil`, not a `Rect`.** It always sits at the rect's top-left, so
+a `Rect` would carry two fields that must equal `rect.left` / `rect.top` and
+could be set not to — the invariant would live in a doc sentence rather than in
+the type. `Component#extent_rect` places it for the two consumers that need
+coordinates (`handle_mouse` hit-testing, `ListDropdown#anchor_to`). Member count
+is a wash; what is bought is that an offset extent cannot be written.
+
+**Consequences.**
+- Four widgets that called `super` now clear only outside the extent, which is
+  strictly less blanking: an unchanged `Checkbox` repaint went from 48 to 22
+  emitted bytes and stopped re-emitting its caption.
+- `Select`'s hand-rolled tail arithmetic is deleted; its `repaint` is the same
+  shape as the other four.
+- Hit-testing stays a per-widget one-liner, because the *action* differs
+  (toggle / open / click) and click-to-focus is deliberately ungated by geometry.
+- **Not** a licence for a parent to consult `extent`. If a container ever wants
+  to, that is the bottom-up channel again and needs its own argument.
+
+## D_final_tree — `children` and `parent` are final; no shadow tree (2026-08-30)
+
+**Decision.** `children`, `parent`, `parent=`, `add_child`, `remove_child` and
+`detach_child` may not be overridden. `Component.verify_final!` resolves each
+one and compares its `owner`, raising `Tuile::Error` from `Component#initialize`
+when a subclass has taken any of them. Checked once per class and memoized.
+
+**Why a runtime check rather than the existing prose.** `D_tree_api` already
+said "never override `children`" and `component_spec` already walked a tree of
+every container kind asserting the array and the pointers agree. Both are
+in-repo guards; neither reaches an *app* subclassing `Component`, which is where
+the mistake is most likely and least visible. The failure mode is nasty and
+silent in a specific way: `attached?` walks the **parent chain** while every
+subtree walk uses **`children`**, so a derived `children` yields a component that
+is attached but never painted, a lifecycle hook fired for the wrong set, and a
+click that never reaches a widget the tree still lists. Nothing raises; the
+widget is just dead.
+
+The check earned itself immediately: `component_spec`'s own `container_with`
+helper built its fixtures with `define_method(:children) { kids }`, i.e. the gem's
+test suite was faking the tree it was asserting about. That is now real
+`add_child` wiring.
+
+**Why the check is at instantiation, not at definition.** A
+`Component.method_added` hook fires at load time, which is nicer, but only sees
+a literal `def` in the subclass body — it misses an override arriving through an
+`include` or a `prepend`. Resolving `instance_method(...).owner` at the first
+`new` catches all four routes uniformly, in three lines, at the cost of firing a
+moment later. It does not stop `instance_variable_set(:@children, …)`, and it
+isn't meant to: the goal is catching the accident, not defeating an adversary.
+
+**Rejected: `Screen` holds a private tree of its own**, authoritative regardless
+of what `parent` / `children` say. Two sources of truth that can drift is
+strictly worse than one that can be lied about, and this is `D_tree_api`'s
+slot-desync rule raised to framework scale — `ScreenPane#popups` needed an
+explicit carve-out and a drift assertion just to duplicate *ordering* for one
+list. Every operation would also have to pick a tree, and the right pick differs
+per operation (paint and focus want the physical one, a named slot wants the
+logical one), so each choice becomes a new bug surface. And it would not even
+fix the case that prompted this: routing is per-component methods, so a shadow
+tree does nothing about a `handle_mouse` that ignores half its children.
+
+**The logical/physical axis is real, and it is served by composition instead.**
+A container whose regions are app-swappable holds a {Component::Slot} per region
+— a logical view implemented *over* the physical tree, never beside it. See
+`D_slots`.
+
+## D_slots — `Slot`, a one-child region; `HasContent` stops meaning "one child" (2026-08-30)
+
+**Decision.** Add `Component::Slot`: a `Component` that includes `HasContent` and
+sizes its occupant to its own rect. A container with several regions gives each
+one a `Slot`, wired once at construction. `HasContent` keeps its implementation
+but is re-scoped to mean *I own exactly one child directly*, and loses
+`handle_mouse` to `Component`.
+
+**The problem.** `HasContent`'s rdoc said "a component with one child tops",
+which `Window` has falsified since it grew a footer — and `Window` paid for it:
+`footer=` was a 20-line hand-copy of `content=` including the notify-last
+ordering rule, plus a `handle_mouse` and a `rect=` patch. That is O(slots)
+boilerplate, each copy a chance to get the order wrong. A three-region dialog
+would have been a third copy, or — worse, and the shape actually proposed — would
+have included `HasContent` and silently mis-routed: `HasContent#handle_mouse`
+forwarded only to `content`, so its buttons would have been unclickable.
+
+**Why a component and not a slot mechanism.** The alternative designed first was
+a declared slot order (`SLOTS = %i[content footer]`) plus a `swap_slot` helper
+computing the insert index as "populated slots declared before me". It works, and
+it makes the ordering rule executable. It was dropped because it answers "why not
+on `Component`, then?" badly: it is a new framework mechanism — a class-constant
+convention, ivar reflection to write the backing store in the right order, a
+doc entry — solving a problem composition already solves. `Slot` adds one small
+class and no new concepts, and the nil case that motivated the whole thing stops
+existing rather than being computed: inside a `Slot` the only insert index is 0.
+
+**Rejected: placeholder components.** Keeping a fixed-arity `@children` by
+parking an inert object in every empty slot. This is a shadow tree at 1/10 scale
+(`D_final_tree`): `children.size` stops meaning what it says, `on_tree` visits
+things that aren't UI, and every generic walk must tolerate ghosts forever — all
+to buy index arithmetic. A `Slot` is not a placeholder: it has a rect, it clears
+it, and it routes.
+
+**Rejected: holder sub-containers built from `Layout`.** Wrapping each region in
+a `Layout::Absolute` needs no framework change at all, which is exactly the tell
+— an app can already do it. It costs a tree level *and* rect plumbing per region,
+and it is a placeholder with geometry.
+
+**An empty slot does not collapse.** It keeps the rect it was assigned and clears
+it, so a dialog with no message shows the hole — consistent with the same dialog
+given an empty message string. Closing the gap is the *parent's* arithmetic (a
+zero extent), which is what top-down layout already demands; detaching the slot
+instead would put the index problem straight back, and neither `Layout#add` nor
+`Box#add` takes an index to re-insert at. `Window` uses the degenerate form: an
+absent footer gets an empty rect, because a `Slot` clears what it is given and
+would otherwise blank the bottom border.
+
+**A slot is transparent, in all three channels.** Not `focusable?`, so the focus
+cascades walk past it; `Component#handle_mouse` descends through it; and
+`on_child_removed` is *forwarded to the parent*, because the default repair moves
+focus to `self` and a slot is inert — no cursor, no keys, nothing to bubble from.
+That last one is not theoretical: it broke `window_spec`'s "repairs focus when a
+focused footer is removed" the moment the footer moved into a slot.
+
+**The cost, paid knowingly.** `Window#children` now always holds the footer slot,
+and a footer's `parent` is that slot rather than the window. Both are honest —
+the region genuinely exists whether or not it is occupied — and one thing gets
+*stronger*: content-then-footer ordering used to depend on `content=` inserting
+at 0 and `footer=` appending, and now holds structurally because the slot is
+wired at construction.
+
+**`HasContent` survives, re-scoped.** A marker mixin with no implementation was
+considered and rejected twice over: the swap dance has to live somewhere or every
+includer hand-writes it again (the duplication this deletes), and a `content=`
+meaning "put this in my Slot" is circular, since `Slot` *is* a `HasContent`. So
+it keeps its body and gains a rule for which shape to use — permanent, integral
+content includes it (a typed field's inner `TextField`, `Slot` itself);
+app-swappable regions hold a `Slot`. It stays a mixin rather than per-class
+accessors so a tree walk can find content via `is_a?(HasContent)`, the same
+reason `HasCaption` is one.
+
+**`handle_mouse` folds into `Component`.** The child-walk existed three times —
+`Layout`, `TabSheet` (verbatim) and, narrowed to one child, `HasContent`, with
+`Window` patching a footer branch on top. It is now the base implementation:
+focus self if focusable, then hand the event to every child whose rect contains
+the point. Widgets that resolve clicks inside their own rect (`List`, `Select`,
+`MenuBar`, `Notification`) already override without `super` and are unaffected.
+One behavior change falls out and is a fix: a click on `Window` chrome now lands
+focus on the window, which `AGENTS.md` has described all along.
+
+**Not folded: `on_focus`.** The obvious symmetry — promote `Layout#on_focus`'s
+first-tab-stop walk to `Component` and delete `HasContent#on_focus` — was
+implemented in design and dropped. Unlike the mouse walk these are not
+duplicates: `HasContent` forwards to *its content*, `Layout` searches for the
+first `tab_stop?` descendant, and they disagree whenever content is focusable but
+not a tab stop (a `Popup` wrapping a `Window`). More decisively, `on_focus` is a
+*public app-facing hook* whose default is deliberately "do nothing"
+(`D_hook_visibility`); giving it default behavior changes what `super` means in
+every app override. The mouse walk carries no such risk — its default already
+does something.
+
+## D_confirm_window — `ConfirmWindow`: the component is the builder; every button dismisses (2026-08-31)
+
+**Decision.** Add `Component::ConfirmWindow < Window`: a caption, a prose
+message and a centered row of `Button`s, opened as a content-measured modal
+popup. The component itself is the builder — `#button(caption, mnemonic:, &action)`
+declares any button set; three class factories (`alert`, `confirm`, `yes_no`)
+cover the common shapes — and **every button closes the dialog**. Book ch7
+("The confirm dialog") carries the user-facing story.
+
+**Callback-only, because blocking is impossible.** The shape everyone reaches
+for first — `JOptionPane`, tkinter `askyesno`, `tty-prompt`'s `yes?`, GTK3
+`dialog.run` — *blocks* and returns the answer. Tuile is single-threaded:
+`Screen#run_event_loop` is `$stdin.raw { event_loop }` with the key thread
+already running, so a value-returning modal would need a nested loop
+re-entering raw mode. This is the first thing a contributor will try to "fix";
+it cannot work here.
+
+**One dismissal channel, N action channels.** A button with a block fires it; a
+button without one is a Cancel. ESC, `q`, an outside click and a Cancel button
+are all one event — `on_dismiss`, fired exactly once and only when no action
+button was chosen. `Overlay#on_close` (which fires on *every* departure) is the
+hook, gated by one bool. This is Vaadin's "ESC triggers the Cancel action"
+minus its `cancelable`/`rejectable` booleans: Ruby can say *absent argument =
+absent button*, which deletes the boolean surface a Java API pays for.
+
+**Every button dismisses, unconditionally — no keep-open knob.** The
+counter-case was hunted and doesn't exist: a dialog staying open after a press
+is either collecting input (excluded below) or chaining — "Copy files" → a
+copy-progress window — and chaining is the callback's job: it opens the *next*
+window. Activation order is **mark chosen → close → fire**: the bool set before
+`close` is what makes `on_close` skip the dismissal, and the block firing after
+close sees clean focus-repair state (the dialog is already out of `@popups`),
+so a callback opening a follow-up popup snapshots the right prior focus and a
+raising block can't strand a half-open dialog.
+
+**The component is the builder — roads not taken.** Prior art: ~20 blocking
+overloads (Swing), setters plus booleans (Vaadin), a builder object (Android),
+flags plus an `addButton` escape hatch (Qt), buttons-as-data with a result
+index (Electron, Turbo Vision, prompt_toolkit), or no dialog component at all
+(Textual). Against `X.new.tap { … }` already being the house idiom, three
+candidates lost:
+
+- *Kwargs only* dies at button 4, and each new knob is a constructor parameter
+  forever. It survives as the **factory** shape, pinned at one or two buttons,
+  where those costs never fire.
+- *Buttons as data + one `case` callback* dissolves once the component is the
+  builder: its one advantage (N buttons with zero API growth) is the
+  mechanism's job now, and it invents a `[symbol, label]` vocabulary next to
+  `Button.new("Save") { save! }`, which already is caption-plus-action.
+- *A separate builder object* is a second class whose only job is to be a
+  half-built dialog, while Tuile components are already mutable.
+
+New capability lands as a `#button` kwarg or a method, never a constructor
+parameter — the exponential growth never starts. And `#button` takes a caption
+plus kwargs plus block, never a prebuilt `Button`: the dialog must restyle the
+caption (the mnemonic underline) and wrap the action (close-then-fire), and
+doing either to a caller's `Button` is spooky mutation.
+
+**Three factories, and no more.** The sets toolkits ship: OK · OK/Cancel ·
+Yes/No · Yes/No/Cancel · Retry/Cancel · Abort/Retry/Ignore. `alert`
+(acknowledge), `confirm` (its labels are kwargs, so it *is* OK/Cancel and
+Delete/Cancel) and `yes_no` (one line over `confirm`) cover them; everything
+further is a label respelling or five lines of the mechanism. Windows'
+six-value `MessageBoxButtons` enum is the tripwire this rule exists to avoid.
+
+**No content slot; `message=` changes kind and stores as given.** The body is
+prose rendered by a `TextView` the dialog owns. Three buys: owning the body is
+what makes scrolling *reachable* (`TextView#handle_key` acts on the key alone,
+so the dialog hand-feeds scroll keys while a button keeps focus); the sizing
+rule has one mode (the dialog always measures content it owns,
+`Notification`-style, rather than "measure unless the caller assigned
+content"); and `StyledString` already covers icons, color and emphasis on a
+TTY. The casualty, priced: exactly the case Vaadin's docs carve out — the
+don't-ask-again checkbox. Everything else people put in a dialog body isn't a
+confirm dialog, and `Popup.new(content: your_layout)` remains the escape
+hatch. **Re-grow rule:** don't-ask-again returns as a named `remember:` seam
+whose state reaches the callback, never as a reopened content slot.
+
+The accessor rule worth keeping: a setter that **normalizes within a kind**
+stores the normalized form (`Label#text`: `String` → `StyledString`, text
+stays text, round-trip pinned); a setter that **changes the kind** stores the
+input and derives the rendering (`String` → a component — handing back the
+`TextView` would hand back machinery). Hence `#message` returns what was
+assigned. The slot occupant is *derived* from the raw value through a single
+write path, so the two stored values cannot disagree (`D_tree_api`'s desync
+rule is about a second copy of tree *structure*). Coercion lives on the
+dialog, not on `Slot#content=`: the right wrapper differs per region — a
+caption-ish line wants a `Label` (ellipsizes), a message wants a `TextView`
+(wraps) — so one `Slot`-level answer would be wrong half the time in this very
+component. A `Component` message is mounted as-is, and the dialog — which may
+measure only content it *owns* (the v0.9.0 re-grow rule's caller-side query,
+here `#measured_size`) — then takes the full half-screen box.
+
+**Mnemonics: MenuBar's shape; `q`, `g`, `G` reserved.** Local sugar over the
+window's own `handle_key` per `D_key_dispatch`'s re-grow rule — never a
+dispatch phase — with the letter underlined in the caption: Tuile has no
+status bar to advertise keys in (`D_status_bar`), so an unadvertised mnemonic
+is a hidden feature. `mnemonic: :auto` (the default) derives the caption's
+first letter and is *silently skipped* when reserved, taken or unusable; an
+explicit letter raises at registration, exactly where `MenuBar#add_item` puts
+every rule that has no sane answer at keypress time. Two-tier on purpose:
+best-effort for a derivation the caller never chose, strict for a promise they
+spelled out. Reserved: `q` — the do-nothing route out of *any* confirm dialog,
+including one that thinks it forces a choice (the user can always Ctrl+C, and
+pretending there is no escape route just trains them to reach for it); `g`/`G`
+— message scrolling; Space — it presses the focused button. The hand-fed
+scroll set (`BODY_SCROLL_KEYS`) is deliberately **not**
+`Keys::UP_ARROWS`/`DOWN_ARROWS`: their vi aliases `j`/`k` stay available as
+mnemonics ("Keep"), and the body still honors them when focused itself.
+Reservation-at-registration is what keeps keypress time free of shadowing
+rules.
+
+**The body is a tab stop; focus opens on the first button.** A plain
+`TextView`, no `tab_stop?`-suppressing subclass: the arrows reach the prose
+either way, but the stop makes overflowing prose *visibly* reachable
+(Shift+Tab) rather than secretly scrollable, and it deletes a nested class.
+`on_focus` lands on the first-declared button — which, since Enter presses the
+focused button, is the default button. A safe-default knob (destructive
+confirms focusing Cancel) can land later as a `#button` kwarg.
+
+**Buttons live inside the window, not in `Window#footer`.** A `Horizontal` as
+the bottom row of the inner `Vertical` (body `Expand[1]`, row `Fixed[1]`,
+spacing 1): the footer paints *over the bottom border row*, and `[ Delete ]`
+glyphs embedded in the border look wrong — the border stays clean chrome.
+Centering rides `cross: Fixed[row width], align: :center`, re-declared on each
+`#button` call, because a `Box` constraint changes only by remove-and-re-add.
+
+**Sizing: measured, capped at half the screen.** The private `MeasuredPopup`
+derives `declared_size` from `#measured_size` on every `reposition`, so a
+message change and a SIGWINCH both re-measure against the current screen — the
+`Overlay` "derived position needs its own `reposition`" rule applied to a
+derived *size*. The cap is `Fraction::HALF`, `Popup`'s own default, so the
+dialog only ever *shrinks below* the default popup box. It re-measures freely
+rather than grow-only like `Notification` — a dialog's text changes far less
+often than a toast's. No floor for now; the risk a floor would hedge (a tiny
+yes/no box going unnoticed over a busy screen) is really a backdrop problem —
+`ideas/modal-backdrop.md`.
+
+**Named `ConfirmWindow`, not `ConfirmDialog`.** It sits in the `*Window`
+family — a `Window` subclass, tiled-or-popup for free — and obeys the
+widget-suffix rule. `ConfirmDialog` is what a searcher will type, but Tuile
+already calls `Popup` "the modal dialog", so that name would imply `< Popup`;
+the rdoc and README say "the confirm dialog" in prose, so the search still
+lands.
+
+**Two seams deliberately not added.** No `header=`: it would be the title, and
+`HasCaption#caption` already is — two accessors for one thing, one storing
+as-given and one coercing, is exactly the wart the store-as-given rule above
+exists to prevent. A *rich* header (an icon beside the text) would come back
+as a region distinct from the title, never as a second name for it. And not a
+`HasValue`: a dialog outcome is not a field value — the same reasoning that
+keeps `ProgressBar` out of the mixin (`D_progress_bar`).
+
+**An alert keeps its OK button** even though ESC/`q`/outside-click already
+close it: Tuile deliberately advertises no quit key (`D_quit_key`,
+`D_status_bar`), so the button *is* the discoverability affordance — and it is
+clickable, which the keys are not.
+
+**Rejected: folding `PickerWindow` in.** A keystroke-addressed, scrollable
+`List` of options with a cursor vs. a short focusable row of buttons with a
+default and a dismissal: one widget with a mode flag would disagree with
+itself on every question that matters — does the cursor roam, is there a
+default, what does ESC mean, does a pick close. What the two share is API
+*shape* — a caption, a set of choices, one callback — not code.
+
+## D_info_window_body — `InfoWindow`: two body presentations, prose wraps and rows don't (2026-08-31)
+
+**Decision.** `InfoWindow` gains `ConfirmWindow`'s body seam — `message=`
+accepting `Component | String | StyledString | nil`, text rendered by a
+wrapping, scrollable `TextView`, the reader returning what was assigned
+(`D_confirm_window`'s store-as-given rule) — and **keeps `lines=`**, rebuilt
+as sugar that mounts a `List` through the same slot (delegating to
+`List#lines=`, so the coerce/split/rstrip semantics stay one implementation).
+The constructor and `.open` take one body positional and dispatch by type:
+`Array` → rows, anything else → `message=`. Last writer wins; after `lines=`,
+`message` reads the mounted `List` back — store-as-given holds, because the
+sugar *assigns a Component*.
+
+**Truncation is a presentation, not a bug.** The itch was that `InfoWindow`
+(born before `TextView` existed) truncated long lines while
+`ConfirmWindow.alert` wrapped them — two "here's some information" paths
+diverging on a long sentence. But the fix is not to kill the `List` body:
+`TextView` has no truncate mode, so a `List` is the only way to show columnar
+output (a file listing, aligned key-value rows) where a wrap destroys the
+alignment. The actual bug was that wrap-vs-truncate was **accidental** —
+decided by which class you reached for. Two named setters on one class make it
+chosen: `message=` is *prose*, `lines=` is *rows*. The docs demote `lines` to
+second billing on purpose — book ch7 and the README row lead with `message=`,
+or everyone keeps reaching for the truncating path out of habit.
+
+**Non-breaking by choice, not necessity.** Neither virtui nor pikuri-tui used
+`InfoWindow` (or `ConfirmWindow`) at all when this shipped, so the breaking
+design first sketched — delete the lines API outright — would have cost
+nothing downstream. `lines=` survives because the rows presentation earns it,
+not for compatibility. Which seam downstream apps actually reach for remains
+worth watching; if `lines=` goes unused for a few releases, *that* is the
+evidence for retiring it. One observable change shipped anyway: a bare
+`InfoWindow.new` now has no body (`content` is `nil`) where it used to mount
+an empty `List`.
+
+**`ConfirmWindow#message=` deliberately does not learn Array→List.** A
+confirm dialog's body is prose by nature; the asymmetry is a decision, not an
+oversight. If a caller ever appears: a `List` of lines is measurable —
+widest-line × row-count, no wrap pass — so it would not fall into
+`#measured_size`'s "Component body ⇒ full half-screen box" hole.
+
+**The coercion is a duplicate, on purpose.** This is copy two of
+`ConfirmWindow`'s `message=` case (per `D_float_field`'s shallow-shell rule —
+fold at four, not two): the classes want different wrappers around the same
+five-line dispatch, and `InfoWindow`'s has no popup to re-measure.
+
+**`InfoWindow` keeps its place next to `ConfirmWindow.alert`:** tiled use, a
+buttonless popup, `declared_size:` control, and the rows presentation — none
+of which the alert offers.
+
+## D_inverse — `Style#inverse`: model SGR 7, don't fake it with colors (2026-08-31)
+
+**Decision.** `StyledString::Style` gains a seventh attribute, `inverse`
+(SGR 7 on / 27 off), plumbed everywhere a style attribute lives: the
+`Data` member, `sgr_to`'s minimal diff, strict `parse`, and a whole-string
+`StyledString#with_inverse` beside `with_bold` / `with_underline`. One
+special ruling rides along: **`under_bg` treats an inverse span as already
+backgrounded** and skips it, exactly like a span with an explicit bg.
+
+**Why an attribute, not a color pair.** The motivating use is the
+inverted focus chip (`[1]-VMs`, LazyVim-mode-segment style). Inverse swaps
+whatever fg/bg are *actually in effect* at the cell — terminal defaults
+included, which `fg:`/`bg:` cannot name — so a chip built with it is
+legible on any terminal palette with zero color decisions. The faked
+version (explicit `fg: :black, bg: <accent>` per theme variant) works but
+re-litigates contrast per theme and still guesses wrong on user-customized
+palettes. That asymmetry — the terminal knows its own default pair, the
+app never does — is the whole case for modeling the attribute.
+
+**The `under_bg` ruling.** `under_bg` fills bg only into spans that have
+none; an inverse span's bg member is nil, but filling it would backfire —
+SGR 7 swaps the effective pair, so the filled tint becomes the chip's
+*glyph* color while its visual background stays the terminal's default fg.
+Skipping keeps an inverted chip looking identical on a plain and a tinted
+panel, which is what "terminal-theme-proof" has to mean. `with_bg` is
+untouched: it is override-all by contract, and a caller explicitly
+assigning a bg to an inverse span gets exactly that (the swap then applies
+to the explicit pair).
+
+**Alternatives rejected.**
+- *The theme-token workaround* (a `fg`+`bg` pair per variant) — see above;
+  it also puts a per-widget contrast decision into every app theme.
+- *Also modeling blink/conceal/dim while in there.* Declined: each
+  attribute costs a code pair in three places plus round-trip and lenient
+  surface, and none has a component waiting. The strict parser keeps
+  raising on them, which is the round-trip contract doing its job — model
+  an attribute when a consumer appears, not for SGR completeness.
+- *Naming it `reverse`* (ECMA-48 says "negative image", terminfo says
+  `rev`). `inverse` is what CSS (`filter`), xterm docs and most modern
+  terminal emulators call it, and `reverse` collides with Ruby's
+  `String#reverse` / `Array#reverse` at the reference site.
+
+## D_background_rgb — `detect` yields a `Result`; the flip re-probes for the RGB (2026-08-31)
+
+**Decision.** `TerminalBackground.detect` returns a `Result(scheme:, color:)`
+instead of a bare `Symbol`, and `Screen#background_color` exposes the color half
+as a `Color` (nil when nothing reported one). It stays current across OS
+appearance flips: `Screen#on_color_scheme` writes `TerminalBackground::QUERY`
+from the event-loop thread, the key thread reads the reply back through
+`Keys.getkey`'s new `\e]` drain, and `EventQueue::BackgroundColorEvent` carries
+it up to `Screen#on_background_color`. A changed color fires
+`Component#on_theme_changed` across the tree, the same fan-out a theme swap uses.
+
+**Why expose it at all.** A theme picks colors to sit *against* the background;
+the borderless-panes idiom (LazyVim's editor-vs-explorer split, virtui's ask)
+derives one *from* it — a secondary pane at ±4–5% luminance, same hue, with the
+primary pane left at the terminal default. That needs the actual RGB, and the
+OSC 11 reply already carried it: `REPLY` captured three components and the
+private `classify` collapsed them to `:light`/`:dark` and dropped the rest. The
+workaround is a fixed near-neutral per variant, which looks right only near the
+background it was tuned on.
+
+**Why a `Result`, not a second entry point.** One OSC 11 exchange yields both
+facts, so one method returns both. `detect` + a sibling `detect_color` would
+mean either a second round trip on a probe that is already timing-constrained,
+or a stashed module ivar that lies the moment `detect` is called twice with
+different IOs — the issue's own second suggestion, declined for that. The
+breaking return type costs exactly one internal caller.
+
+**Why the live re-probe, rather than a documented startup snapshot.** Mode 2031
+reports light/dark and no RGB, so an unrefreshed value survives a flip pointing
+at the *old* background — a dark-derived tint on a now-light terminal, silently,
+in the one situation the theme machinery otherwise handles perfectly. The
+startup-only timing constraint (the reply lands on stdin, which the key thread
+owns once the loop runs) dissolves once the key thread is the one reading it.
+
+Three placements make that safe, and each is load-bearing:
+
+- **The query is written from the event-loop thread**, not the key thread. That
+  thread also owns `emit`, so the query's bytes can never land inside a frame's
+  synchronized-output batch. A key thread writing its own query would race
+  every repaint.
+- **The reply is drained a byte at a time** in `getkey`, like `read_paste` and
+  for a sharper reason: an OSC reply may end in ST, which *is* `\e\\` — a
+  gulping read would swallow the terminator plus whatever was typed behind it.
+  `\e]` can't collide with a keyboard sequence, so the drain eats no real key.
+- **`print` now flushes.** It never did, and every existing caller got away with
+  it because a frame's `emit` flushed moments later. A *query* whose reply the
+  app is waiting on cannot rely on that.
+
+**The stale value is kept, never blanked.** Nil-ing `@background_color` on the
+flip and refilling it on the reply is the honest-looking option and is wrong: a
+terminal that reports 2031 flips but not OSC 11 would lose, permanently, the
+color it gave us at startup. Holding the old value costs one frame of a slightly
+wrong tint on terminals that *do* answer, and costs nothing on those that don't.
+
+**Why `on_theme_changed` and not a new callback.** The hook's contract is
+"rebuild the colors you derived from the theme"; a background-derived tint is one
+of those, and its inputs just moved. A dedicated `on_background_color_changed=`
+would be a second channel firing microseconds after the first, for an app that
+must handle both identically. The cost is one extra full repaint per OS
+appearance flip, which is a rare event with a full repaint already in it.
+
+**Alternatives rejected.**
+- *A module-level `TerminalBackground.background_color` accessor* — mutable
+  module state, stale by construction. See above.
+- *Tuile computing the tint* (a `Theme#tinted` or a `Color#lighten`). Out of
+  scope: how far to step, in which direction, and whether to step at all is the
+  app's design decision, and Tuile has no component that wants it. Exposing the
+  fact is the framework's job.
+- *Deriving the scheme from the re-probe's RGB* instead of trusting the 2031
+  report. They agree unless the terminal is buggy, and the report is the thing
+  that actually said "the user flipped their OS appearance" — so the event
+  carries the color alone.
+- *A `FakeScreen` pinning a plausible dark RGB* rather than nil. Nil is what a
+  non-answering terminal reports, which is the branch app code most needs
+  exercised; a spec that wants a color assigns one through
+  `FakeScreen#background_color=`, which takes the same path a real reply does.
+
+## D_color_depth — Detect the depth; degrade RGB at the wire (2026-08-31)
+
+**Decision.** `ColorDepth.detect` reads the terminal's color depth from the
+environment (`:truecolor` / `:palette256` / `:ansi16`), `Screen#color_depth`
+holds it for the session, `Color#quantize(depth)` maps a color to the nearest
+one that depth can show, and `Buffer#flush` applies that to every color on its
+way to the wire. Upstream issue #8.
+
+**Why it became necessary.** `Color#sgr_codes` emits `48;2;R;G;B` for every RGB
+color unconditionally, which was fine while RGB only ever came from a
+declaration site — a human picking a theme constant for a terminal they were
+looking at. `D_background_rgb` changed that: an app can now *read* the
+background and *derive* a color from it (virtui's borderless-pane tint steps the
+reported background toward its own pole), so Tuile hands out an RGB value the
+app has no safe way to write back out. Under a 256-color terminal, or tmux
+without `terminal-features "*:RGB"`, that computed `48;2;…` is mangled or
+silently approximated.
+
+**Why the downgrade is automatic, and why it lives in `Buffer#flush`.** The
+deciding argument is that *not all RGB has a call site to opt in at*. RGB enters
+an app three ways: **declared** (a `Color.hex` theme token), **computed** (the
+tint), and **parsed** — `StyledString.parse` ingests ANSI produced by other
+programs, e.g. a `Component::LogTextView` fed a tool's colored output. Parsed
+colors arrive as *data*, with no declaration site an app could quantize at, and
+`StyledString` must stay depth-unaware (it is a frozen value type with a
+`parse(to_ansi(x)) == x` round-trip and no `Screen` dependency — the same rule
+that keeps it theme-unaware). Only a choke point on the wire catches that case.
+
+`flush` is that choke point: it is where logical cells become bytes, the same
+role `draw_text` plays for backgrounds. Quantization happens *before* the
+`Style#sgr_to` diff, so two RGBs landing on one palette cell emit a single SGR
+instead of two. The "but then `sgr_codes` is dishonest" objection dissolves at
+this placement — `Color` still emits exactly what it was given, and `flush`
+already doesn't emit what you wrote (it skips unchanged cells and wraps frames
+in sync batches). Adapting a logical frame to a physical terminal *is* the
+buffer's job. Nor does this reopen the framework's allergy to automatic
+channels: the deleted ones (`content_size`, `keyboard_hint`) were semantic
+queries the framework made *of components*; this consults nobody, adds no
+`Component` API, and is one field on `Buffer`.
+
+**Roads not taken:**
+
+- *Opt-in `quantize` only*, leaving apps to call it. Serves the computed case
+  and nothing else — see the parsed case above. `quantize` stays public anyway,
+  for an app that wants to *know* what a color becomes on the wire (checking a
+  computed tint still contrasts with the background after both round to palette
+  cells) without changing what it stores.
+- *Pre-quantizing at theme definition or at tint derivation.* Redundant (flush
+  catches those anyway) and actively harmful: it bakes depth into stored state,
+  the cache-in-an-ivar failure the theme and `bg_color` rules forbid. A stored
+  `Color.palette(237)` has forgotten it was `#3a3a3a`, so any later contrast
+  check or derivation works from the lossy copy. `ThemeDef.default` is built at
+  load time anyway, before a `Screen` exists to supply a depth. The model is:
+  **logical layer always truecolor, wire layer always terminal-native, one
+  conversion at the boundary.**
+- *Raising at render on an unrepresentable color*, leaving apps to supply only
+  representable ones. The fail-fast instinct matches the house "raise at
+  registration, not gate at runtime" pattern, but that pattern works because it
+  raises *at the write site, deterministically, on the developer's machine*.
+  This inverts both: it fires at the read site, far from the assignment, and
+  only on the *end user's* terminal — a developer's truecolor terminal and
+  `FakeScreen`'s pinned `:truecolor` never see it, so it ships and crashes on
+  tmux. Breaks-at-a-distance and silent-under-test, by design. It also converts
+  "coarser shade" — which the terminal already approximates on its own — into
+  "app dies mid-repaint", and no peer framework does it (Rich, Textual, tcell,
+  chalk and notcurses all degrade).
+- *A keyed cache — memo or LRU — in front of `quantize`.* Measured and
+  rejected (`benchmark/quantize.rb`, 1M calls, ruby 3.3.8): compute ~360
+  ns/call on both workloads; an unbounded memo 158 ns (typical) / 217 ns
+  (gradient, having grown to 100k entries); a 256-entry LRU 183 ns (typical)
+  but **648 ns** on the gradient — 1.8× *slower* than just computing, because
+  every miss pays lookup + compute + eviction. The bounded cache only wins the
+  workload that needed no help, and the unbounded one is keyed on a 16.7M-entry
+  input space with the parsed-ANSI case as its adversary.
+  `Buffer::WIDTH_CACHE` is unbounded for reasons that don't transfer: distinct
+  graphemes are bounded by fonts and languages, and each avoided gem call costs
+  ~20×. What is exploited instead is the *output* space — 240 palette cells and
+  16 names — so frozen tables let `quantize` be pure arithmetic returning a
+  shared instance, allocating nothing.
+
+- *A `Screen#color_depth=` setter.* Detection runs once and the depth cannot
+  change mid-session, `ColorDepth::OVERRIDE_ENV` already covers a terminal that
+  misreports, and a setter drags in a real bug: `Buffer::Cell#set` flips dirty
+  only on a *style* change, but a depth change alters the *bytes* an unchanged
+  style emits — the minimal diff has nothing to notice, so the setter would owe
+  a buffer-wide `mark_all_dirty`. Deleting the setter deletes the wrinkle.
+- *Consulting terminfo* (`RGB`, `colors#0x1000000`) as issue #8 suggested. Tuile
+  has no terminfo access — tty-screen does geometry, not capabilities — so this
+  means shelling out to `tput`/`infocmp` at every startup. The env ladder plus
+  the override covers the real matrix, and its failure mode is *conservative*:
+  tmux and ssh (which drops `COLORTERM`, not being in the default `SendEnv`
+  set) under-report, which renders coarser but never mangled.
+- *Two-stepping RGB → 256 → 16* under `:ansi16`, reusing the palette quantizer.
+  Compounds the rounding: `rgb(0, 0, 195)` is nearer bright blue than blue, but
+  rounds to cube cell 19 `(0,0,175)` first and then picks blue off *that*.
+  Direct nearest-of-16 costs one more table and is pinned by a spec.
+- *A "needs translation" predicate or form sugar beside `quantize`.* The need
+  is a function of *(form, depth)* — a bare `from_palette?` is ambiguous
+  between the 256 and 16 targets, and `full_rgb?` misses the palette→16 case —
+  which is exactly the case analysis `quantize` already performs, so a separate
+  `needs_quantize?(depth)` would restate it as a boolean and drift from it.
+  Identity-return makes the predicate free instead:
+  `color.quantize(depth).equal?(color)`. Form introspection (`named?` /
+  `palette?` / `rgb?`, matching the factory names) is likewise left out until
+  an app asks — `color.value` answers it. Same shape one level up: a
+  `Screen#truecolor?` boolean instead of the three-valued symbol would leave
+  `:ansi16` inexpressible.
+- *A perceptual distance metric.* Plain squared-Euclidean RGB. The place a
+  perceptual weight would show is the cube-vs-grey-ramp tiebreak on near-greys
+  — exactly what a background-derived tint produces — so `color_spec` pins a
+  real stepped tint (`rgb(30,30,34)` → `palette(234)`), and the metric gets
+  revisited only if that cell ever looks wrong on an actual screen.
+
+**One memo *is* needed, and finding that took measuring the right thing.**
+`Buffer#quantized_style` runs per dirty **cell**, not per style transition —
+easy to get wrong from the design sketch, since `sgr_to` is what runs per
+transition. So a full-screen repaint of RGB-styled content paid the arithmetic
+8000 times for what was one span, and measured **51 ms against 15 ms** at
+`:truecolor`: a 3.4× regression landing squarely on the app the feature exists
+for, whose panes are painted in a computed tint. The fix is a *one-slot* memo —
+remember the last `(style → quantized style)` answer and compare by identity,
+sound because a `Style` is frozen. That collapses the work back onto genuine
+transitions and makes depth nearly free (10.7 ms vs 9.9 ms at `:truecolor`;
+`:ansi16` likewise). It is not the rejected cache in miniature: no key space, no
+eviction, two ivars, and it exploits run locality within a row rather than value
+recurrence across a session. Two supporting micro-optimizations in
+`nearest_palette` came out of the same pass — destructuring rather than
+splatting the triple, and `x * x` rather than `x**2`, together ~2× — which is
+why that method is written flat instead of tidy.
+
+**A known-lossy corner, accepted.** `:ansi16` matching uses xterm's default RGBs
+for the 16 named colors, which a terminal's own scheme may redefine — the one
+mapping here that can be honestly wrong. It is documented on `Color#quantize`,
+and the result is a *named* color (SGR `30..37`/`90..97`), so the user's scheme
+still decides what is finally drawn. `TERM=linux` is about the only consumer.