perf(j2k-native): skip tiles outside the output region before workspace build

[CarlAllenn]

Problem

Region decode iterates every tile in the codestream, running build::build (decomposition/precinct/code-block graphs plus coefficient-buffer resize/zero) and segment::parse per tile before the ROI plan can skip any entropy work. Per-request cost therefore scales with the image's total tile count even for a tiny ROI.

Measured on a 15000×11000 RGB codestream with 1024 px tiles (165 tiles, 7 decomposition levels), M1 Pro, serial mode, warm decoder:

request	before	after
64 px ROI, one tile	63 ms	4 ms
512 px ROI, one tile	90 ms	27 ms
same ROI on a 35-tile codestream	40 ms	27 ms

(A sampling profile attributed the fixed cost to Vec::resize/__bzero under j2c::build::build and to tag-tree/segment parsing, repeated for all 165 tiles per decode.) After the change, per-request cost is proportional to the ROI, not the codestream.

Change

In the decode tile loop, skip a tile when its rect cannot intersect the output region. The intersection test maps the region onto the reference grid the same way RoiPlan::build does — including both the component subsampling factor and the reduced-resolution shrink factor — and rounds the tile rect outward, so any rounding disagreement decodes a boundary tile rather than skipping it. Skipped tiles contribute no samples to the stored output, so decoded pixels are unchanged.

Tests

Adds scaled_region_decode_on_multi_tile_codestream_matches_scaled_whole_decode_crop: a multi-tile codestream with partial edge tiles, regions interior / straddling / far-corner, at every Downscale variant, compared against the scaled whole-image decode. This covers a gap the existing suite has (multi-tile × scaled region): an earlier draft of this change ignored the resolution shrink factor and every existing test still passed; this test fails on that draft and passes on the final change.

Validation

cargo xtask fmt / clippy / clippy-strict / test / release-integrity / repo-lint on macOS arm64 and Ubuntu x86-64. The only test failure in either environment is the pre-existing unsafe-audit gap on main (#64), which this branch doesn't touch.

🤖 Generated with Claude Code

[jcwal1516]

This is a real performance problem, and the patch merges cleanly onto current main. I ran the native and facade decode suites successfully. Before merging, I’d like focused coverage for nonzero image/tile origins and component-subsampled codestreams, since an incorrect early-rejection calculation could silently omit contributing tiles. Please also consider sharing or consolidating the overlap calculation with the existing direct-plan tile-intersection logic so the two paths cannot drift. The optimization itself looks worthwhile.

[jcwal1516]

The extra coverage and shared overlap helper were an easy fix, so I went ahead and added them and merged this. Thanks for the contribution!