render, types, ci: resolution-time dp with per-canvas density, bounded GLES image cache, clippy gate, backend capability matrix

Length::dp no longer collapses to absolute pixels at construction: the design value travels in a new LengthBase::Dp variant and the density multiplication happens when the length is resolved. Previously dp( n ) baked in whatever density() returned at view-build time, so correctness across output changes depended on the view being rebuilt after set_density and in that order; now a density change is picked up by the very next paint with no reconstruction. Length::resolve keeps its signature (process density), and the new Length::resolve_with_density takes an explicit factor. dp becomes const in the bargain.
Density also becomes overridable per canvas, the first step towards surface-local responsive state. SoftwareCanvas and GlesCanvas carry a density: Option<f32> analogous to the layout_viewport introduced for sub-canvas fluid resolution: None means "use the process global", Canvas::set_density pins a local factor, and sub-canvases inherit it. All canvas-routed resolution honours it — geom_px / font_px for stock-widget design pixels, and the new Canvas::resolve_geom / resolve_font for explicit Length values, which every widget now uses in place of the raw l.resolve( canvas.viewport_layout(), EM ) pattern (row, column, wrap_grid, spacer, container, separator, button, text, rich_text, text_edit, list_item, vslider, image, and the container draw path). Overlay sizing keeps resolving against the main surface with the global density, which is what it describes. New tests cover explicit-density resolution, resolution-time application, the local-over-global override and sub-canvas inheritance.
The GLES image texture cache is now bounded. It was content-keyed but unbounded and never evicted, so a stream of distinct buffers — a photo carousel, video thumbnails — grew GPU memory for the lifetime of the canvas. The cache now tracks an estimated byte total (RGBA8, w × h × 4) against a 32 MiB budget and evicts least-recently-drawn textures on insert; the most recent entry is never evicted, so a single texture larger than the whole budget still draws and simply owns the cache until replaced. Drop-time cleanup is unchanged: drain deletes whatever the map holds.
CI gains a Clippy step (workspace, all targets, test-support, -D warnings) sharing the build cache of the test job, with make clippy mirroring the invocation locally and CONTRIBUTING listing it. Run make clippy locally before pushing the first time — the gate has not seen the tree yet and pre-existing lints will fail CI until addressed.
docs/backends.md formalises the software/GLES capability matrix that was previously scattered across per-method rustdoc: parity set (fills, strokes, text, images, paths, path clips), graceful degradations on software (flat-fill gradients, no shadows, no backdrop blur, hard bottom edge), GPU-only features (external textures), the shared Oklab-fallback limitation, and the cross-backend blit panic. Linked from README, onboarding and architecture's known-gaps list, which now states the parity gaps explicitly. The dp/density prose in architecture.md, lib.rs and the Length rustdoc is updated for resolution-time semantics and the per-canvas override.

docs/architecture.md

@@ -193,11 +193,11 @@ Every size in a widget tree is a `Length`, resolved to concrete pixels at layout
 `ltk` offers two adaptation strategies, and both live in the same `Length` type so an app can mix them per value:
 
 - **Fluid** (`Length::fluid(n)`, and the raw `vmin` / `vmax` / `vw` / `vh` / `orient` units): surface-proportional. `fluid(n)` reads a single design pixel `n` as `vmin(n / fluid_reference() * 100).clamp(n * FLUID_MIN, n * FLUID_MAX)` — at a surface whose short side equals the reference (412 px by default) it is exactly `n`, and it scales with the short side elsewhere, auto-clamped to `[0.7n, 1.5n]`. This tracks the width in portrait and the height in landscape, because the short side *is* the width in portrait and the height in landscape. `orient(portrait, landscape)` is the escape hatch for a different percentage per orientation.
-- **Physical** (`Length::dp(n)`): constant physical size. `dp(n)` is `n × density()`, where `density()` is a process-wide factor (default `1.0`, typically set from the output DPI via `set_density`). It does not scale with the surface, only with pixel density — the mainstream HiDPI `dp`.
+- **Physical** (`Length::dp(n)`): constant physical size. `dp(n)` resolves to `n ×` the pixel density (default `1.0`, typically set from the output DPI via `set_density`). It does not scale with the surface, only with pixel density — the mainstream HiDPI `dp`. The multiplication happens at **resolution time**, not when the `Length` is constructed: a `dp` value carries its design pixels, so a density change takes effect on the next paint without rebuilding the view's lengths.
 
 Stock widgets do not hard-code either strategy. Each carries a design pixel per dimension (e.g. `button` height 48, font 16) and resolves it through the process-wide `WidgetScaling` mode: `Length::widget(n)` returns `fluid(n)` under `WidgetScaling::Fluid` (the default) or `dp(n)` under `WidgetScaling::Physical`. `set_widget_scaling(mode)` flips it once for the whole app. An explicit `Length` on an individual widget (`button.height(...)`, `text_edit.height(...)`, `font_size(...)`) bypasses the mode entirely — the mode only decides the meaning of the *default* design pixels, never an override the app wrote on purpose.
 
-Both `density()` and `widget_scaling()` are process globals read during layout; set them at startup (or, for density, whenever the surface moves to an output with a different DPI). Because they are global, ltk's own test suite serialises the tests that touch them.
+Both `density()` and `widget_scaling()` are process globals read during layout; set them at startup (or, for density, whenever the surface moves to an output with a different DPI). Because they are global, ltk's own test suite serialises the tests that touch them. Density is also overridable **per canvas**: `Canvas::set_density` pins a canvas (and every sub-canvas derived from it) to its own factor, and all canvas-routed resolution — `geom_px` / `font_px` for stock-widget design pixels, `Canvas::resolve_geom` / `resolve_font` for explicit `Length` values — uses the local density when one is pinned and the process global otherwise. This is the hook for a surface sitting on an output whose DPI differs from the one the global was derived from (an overlay on a second monitor, an embedder with several `UiSurface`s).
 
 `Length` adapts *sizes* to the orientation; to adapt the *structure* of a layout (a row of panels in landscape, the same panels stacked in portrait), branch the view on `ltk::orientation()`. The runtime records the main surface's physical dimensions on every configure (also readable as `ltk::viewport_size()`) and rebuilds the view after each resize, so a `match ltk::orientation() { Landscape => row()…, Portrait => column()… }` follows the window live. The portrait/landscape rule matches `Length::orient` (a square surface counts as portrait). `examples/clip_path.rs` shows the pattern.
 
@@ -322,3 +322,5 @@ Downstream consumers shipping into regulated environments (EN 301 549, WCAG 2.1
 **xdg-activation-v1 — wired in.** Both directions work: a token found in `$XDG_ACTIVATION_TOKEN` at startup is used to activate the app's own window once it maps (so an external launcher can raise an ltk window with focus), and an app that spawns children requests fresh tokens through `App::take_activation_requests` and receives them via `App::on_activation_token` to place in the child's environment.
 
 **Fractional scale — deferred.** `wp_fractional_scale_v1` (so 125 % / 150 % outputs render natively instead of via compositor downscale) remains tracked as upcoming protocol work.
+
+**Software/GLES parity gaps — see [`docs/backends.md`](./backends.md).** The software backend renders gradients as a flat fill from the first stop, skips outer and inset shadows and backdrop blur, and hard-cuts the bottom-edge fade; `oklab` gradient interpolation falls back to linear-light on both backends. The capability matrix is the canonical per-feature table and must be updated in the same patch that closes any of these gaps.