evict layer backdrop textures on resize and destroy to prevent VRAM leak

[sophietheopossum]

The layer-surface counterpart to 1ac0829 (#59), which fixed this leak class on the window side; both new helpers land beneath its purge_shared_effect_pipeline_caches_for_window and are named to match.

layer_backdrop_cache is insert-only: 42 references across tty.rs, winit.rs and state.rs, not one of them a remove, retain, clear or drain, and because the key ends in the effect rect's size (__layer_background_effect_DP-1_<layer_id>_0_1520x471), a resize mints a fresh full-size texture for every size a layer has ever had. The texture is allocated in physical pixels while the key carries the logical rect, so that key costs 2.7 MiB at scale 1.

Scope: only configs that put a backdrop effect on a layer surface, a per-layer behind: effect, or background_effect and a client declaring a blur region through ext-background-effect-v1. A stock config never writes an entry, and a plain backdrop blur on a Top/Overlay layer takes the framebuffer path instead, so in practice it's behind: effects and Background/Bottom layers.

Nothing reclaims those entries today. layer_destroyed doesn't touch the map; the hot-reload reset at state.rs:2481 clears five sibling effect caches and skips this one; refresh_layer_effects_for_output sweeps configured_layer_effects and layer_effect_cache for dead layer ids and skips this one; and the 128-entry pipeline LRU frees the intermediates while the final texture stays pinned by the clone stored here.

What changed

• evict_stale_backdrop_sizes(cache, current_key) at the four insert sites (two in tty.rs, two in winit.rs): splits the key at its last _ and keeps only the current size for that layer and variant. The current key is exempt so the insert can still read back its element Id and commit counter. It runs on the insert path only: the cache-hit branch returns before it, though an Always-invalidation effect takes that path every frame; the retain is a prefix compare per entry over a map bounded by outputs × layers-with-backdrop × variants, inside a function already doing a full-size capture and N blur passes.
• purge_backdrop_cache_for_layer from layer_destroyed, clearing a departed layer across every output. On an abrupt client exit wl_surface().client() is already None, so layer_runtime_id degrades to unknown-client:<id> and the purge matches nothing.
• retain_backdrop_cache_for_live_layers closes that gap, called from both refresh_layer_effects_for_output paths beside the existing layer_effect_cache sweep: it drops any entry whose layer is not in the live set, so it needs no event to fire and also covers a close the compositor missed. It cannot reuse retain_effect_texture_cache_for_live_ids: that tests {id}@ as a key prefix, while these keys carry the id between underscores after the output name, so handing it this map would match nothing and empty the cache on every refresh.
• Three of the four sites keyed on a bare wl_surface().id().protocol_id(), unique only within a client, rather than layer_runtime_id() ({client_id}:{protocol_id}); the purge searches for _{layer_runtime_id}_, so it cannot match a key built from a bare protocol id. The correct binding was already in scope at all three: tty.rs's top-layer and winit.rs's lower-layer sites each ignored one earlier in the same function, and winit.rs's top-layer site shadowed one, so that fix is a deleted line, which also re-keys that function's three sub-element rect keys from the protocol id to the runtime id (reads and writes move together, so it's inert, but it's why one deletion touches more than stable_key). On its own the collision reads as cache thrash plus two render elements sharing a damage-tracker Id and commit counter, since the signature check, which hashes the effect rect and capture geometry, gates the actual texture reuse. Found by inspection, not reproduced.

Testing. Both new paths fire in a real session here (Intel xe iGPU):

purged destroyed layer backdrop textures layer_id="...:59" removed=1
evicted resized layer backdrop textures current_key="__layer_background_effect_DP-1_...:36_0_1520x471" removed=1

cargo check -p shoji_wm --bins is clean against main at 0d4eab1 itself: I built this branch in a fresh worktree on that base, zero warnings. cargo test -p shoji_wm --bins I have only run on my own tree: 186 passed, 4 failed, all ssd::evaluator::tests::workspace_* kinetic-scroll assertions that fail the same way with this patch stashed, so they are not caused by the change; my tree diverges from main inside ssd/, so I can't say whether they also fail for you. The session logs above were captured on my tree too, though every region this patch touches is identical in both, and the branch is one commit on 0d4eab1, so it applies as a fast-forward.

Two commits: the eviction and the keying fix, then the liveness sweep on top. The keying fix rides with the first because it is the same expression built wrong, happy to lift it into its own commit if you'd prefer that.