perf: the interactive path - #54

This review was written by Claude (Fable 5), acting on Astrid's behalf.

The core algorithm work here is strong: the running-sum box blur, parallel lens blur/edges/warp/convolve/inpaint, disc-morphology decomposition, incremental seam-carve energy, scanline lasso fill, Felzenszwalb–Huttenlocher EDT, LRU tile caps and one-undo quick-select all check out as output-equivalent, and the equivalence tests inline the exact replaced implementations as references with real boundary cases. What blocks the merge is the app-shell glue around them:

1. Selection + generator filters regress (workspace.rs, filter_region_with_context ~4711): with a selection active, the region is now intersected with content_bounds().inflated(pad) (pad = 0 on the dialog path). Render ▸ Clouds on a new empty layer creates content — content_bounds() is EMPTY, the intersection is empty, and begin_filter_preview refuses with "Nothing to filter" where main fills the selection. A selection extending past existing content similarly gets its clouds cropped.

2. The headline "filters read through the selection edge" fix is dead plumbing (apply_filter ~4925 / begin_filter_preview ~4860): filter_region_with_context with context > 0 is unreachable — parameterless filters pass 0, and the dialog's Apply reuses preview.region, which begin_filter_preview computed ungrown via filter_region. So the whole FilterPlugin::context() surface (plugin-api, the blur impls, the simple_filter! heuristic) never grows a buffer on any interactive path, and the selection-edge band the PR body claims to fix is still there. The context_tests only assert that context() returns numbers. The growth needs to happen in begin_filter_preview.

3. Prefetch permanently disabled on large documents (prefetch_tick ~5464): if self.cache.is_full() { self.prefetch_queue.clear(); return false; } — steady state after eviction is bytes >= budget, so once a 16k×16k document fills the 256 MiB budget, every rebuilt nearest-first queue for a new viewport is discarded each tick, including the mid-gesture warming main's own comments call out as what makes the settle frame land instantly. The comment's premise ("what is already in is what the viewport actually needs") is wrong the moment the viewport moves. Prefetch the near ring and let LRU evict distant tiles.

Smaller items:

• fx_blur.wgsl:4-7 still claims the shader mirrors box_pass "tap for tap… no running total whose float error would drift" — now false; update the comment and ideally add a long-dimension parity case (drift grows with row length; the suite only tests 64×48).
• horizontal_morph's "monotonic deque, O(1) per pixel" is a Vec with remove(0) — O(window) per removal; make it a VecDeque.
• MAX_FILTER_CONTEXT was inserted between AUTOSAVE_SECS and its doc comment (workspace.rs ~87), and box_pass has a stale stacked doc comment — same splice pattern as sibling PRs.
• simple_filter! treating any "amount" param as spatial reach conflates intensity with displacement (Add Noise grows by up to 256 px once item 2 is fixed).
• cost_does_not_grow_with_the_radius asserts wall-clock ratios — a CI flake hazard that mostly tests the new code against itself.
• The cited perf numbers (205 ms → …, 5.62 s, 4.43 s) have no harness behind them despite the repo's fxbench/afbench precedent — plausible given the complexity changes, but worth pinning.

Fix the three majors and this is a clear merge — the algorithmic core is the best-engineered part of this PR series.

This follow-up review was written by Claude (Fable 5), acting on Astrid's behalf.

Most of the round landed well: the generator/selection regression is properly fixed (selection.bounds().inflated(pad) with no content intersection — Clouds into a selection on an empty layer works), the context plumbing is now live end-to-end on the dialog path with consistent preview/commit blit math, the wgsl comment is honest with a 4096-row parity case, horizontal_morph is a real VecDeque, the wall-clock test became a semantic one, and the benches exist. Good work.

One new major keeps this from merging — the prefetch fix trades "permanently dead" for "actively evicts the viewport" on the same large documents:

prefetch_tick no longer stops when full (good), but three facts combine badly (workspace.rs:65, crates/compositor/src/lib.rs:722,840):

• PREFETCH_TILE_BUDGET = 2048 tiles ≈ 0.5–1 GiB, twice the cache's ~1024-tile capacity (256 MiB / 256 KiB per tile) — the queue's own comment says so.
• TileCache::prewarm filters to missing tiles and never refreshes resident entries' touched stamps; paints are served from display_tiles first, which skips cache.get — so nothing re-touches visible tiles while a drain runs.
• Each insert past capacity evicts the minimum-stamp entry, and display_tile's retain drops the display copies of whatever the composited cache evicts.

So on a 16k×16k document a full drain first evicts the viewport, then the near ring, then its own earliest (nearest) inserts; after ~8 s the resident set is the farthest ~1024 queued tiles, the next paint recomposites the whole viewport — exactly the stall prefetch exists to prevent — then rebuilds the queue and loops forever. The new comment's "LRU evicts the distant tiles instead" is inverted. Any one of these fixes it: cap the queue at (cache tile capacity − visible count); have the per-paint visible prewarm touch resident entries; or stop the drain when the next candidate is farther than the nearest would-be evictee.

Two smaller residuals, take or leave:

• The Filter Gallery path (show_filter_gallery → plain begin_filter_preview(), workspace.rs:2493) still previews and commits with an ungrown region, so the selection-edge band this PR fixes for the dialog persists for the same filters applied via the gallery.
• The dialog reach heuristic (workspace.rs:5276) includes "amount", which simple_filter!::context() deliberately dropped one commit earlier — Add Noise previews grow by up to 256 px for a filter that reads nothing. Perf-only, but the two lists should agree.

Also still open from last round: the unbounded failed.join(", ") export status string, and no end-to-end test asserting the region growth actually reaches a filter (context_tests still only check context() numbers). Fix the prefetch inversion and this merges.