perf(imaging): route morphology through Leptonica's DWA kernels, exactly

[P4suta]

Stacked on #32. The main course of the despeckle round.

Why

The cleaner's two heaviest morphology ops ran on Leptonica's generic rasterop bricks: the 43×43 isolated-dust dilation (37.5ms/page — the single hottest op in the committed baseline) and the 7×7 fill-holes opening (12.2ms). Leptonica ships word-accelerated DWA kernels for both; they were unbound.

The trap the equality sweep caught

Binding them surfaced exactly what the planned pixel-identity gate existed for: pixDilateBrickDwa silently diverges from the generic brick for sel sizes missing from the generated table — empirically every prime above 15, including the production 43 — while pixOpenBrickDwa is exact at every size and dilate is exact up to 15. (The composite pix*CompBrickDwa variants diverge too; diagnosed per-size in-container, table in the test rationale.)

The shipped routing (exact by construction)

• dilated(): single DWA pass ≤ 15; larger sizes composed from safe-size DWA passes — exact by Minkowski sum (brick(a) ⊕ brick(b) = brick(a+b−1); per-pass clipping changes nothing inside the image rectangle because L∞ paths between in-bounds points stay in bounds), and version-robust (only the always-present small sels are used). 43×43 = three passes.
• opened(): DWA ≤ 15 (production is 7×7), generic beyond (an opening does not compose).
• Generic variants stay as package-private oracles; PixTest pins pixel-identity across radii 0–31 (sel 1–63) on border-touching ink + degenerate (tiny/all-black/all-white) pages.
• Rider: pixCountPixels reuses one process-lifetime popcount table (hygiene, <0.1%).

Measured

metric	before	after
dilate 43×43	37.5ms	14.1ms (−62%)
open 7×7	12.2ms	4.0ms (−67%)
clean() pipeline path	139.5ms	107.6ms (−38.5% cumulative vs the #31 baseline)
despeckle stage (-j8)	9.57s	5.47s (−43%) — bandwidth relief compounds across workers
conv (-j8)	13.61s	9.51s (−30%; −34% vs the original #28 baseline)

selectBySize is now ~70% of the remaining clean() — the measured gate for the follow-up selection restructuring (De Morgan flip + single-labeling fusion) is met.

Verification

• Full ./gradlew check green; ArchUnit FFM confinement untouched (new bindings in Leptonica, routing in Pix).
• The golden cleaner tests pass unchanged (they exercise both DWA paths).
• Rollback = revert the dilated/opened internals (~10 lines); bindings and sweep stay harmless.

🤖 Generated with Claude Code

Re-file of #33 (same commit, cherry-picked onto main).