[darkclad]

Supersedes #487. It carries that PR's first three commits unchanged and addresses its review: the load_tile() commit is reworded, the trailing close() commit is dropped, and the contour fix and the test the review asked for are added.

What changes

The out-of-core 2D pass streams an oversized ROI's pixels tile by tile and derives each pixel's slide coordinates from the tile index. It computed the tile's column as tileIdx / num_tiles_hor, the same expression as the tile's row, so a pixel's x was right only in a tile on the grid's diagonal, and the ROI's pixel cloud had wrong x coordinates anywhere else. This is the defect #483 mentions; with #483 it surfaces as an error instead of wrong values.

Once pixels reach the right coordinates, a ROI that spans tiles reaches the out-of-core contour, which made its result depend on how the file is tiled. That is fixed here too, since this PR is what makes it reachable.

One commit per change:

1. The column is tileIdx % num_tiles_hor, matching how the image loaders decompose a tile index (phase3.cpp).
2. Rows are bounded by the vertical tile count and columns by the horizontal one, and a tile is indexed row * ntHor + col, as in the other tile scans. The old bounds and stride were transposed but cancelled out, so no tile was skipped.
3. The inner row/col no longer shadow the loop variables.
4. load_tile()'s status is checked, as in the other tile scans. This reports nothing today: load_tile(size_t) refuses only an index outside the grid, which this scan never builds, and a tile the TIFF reader fails on arrives zero-filled with a true status, so its pixels are still dropped without a message.
5. The scan places pixels by the loop's own row and col instead of decomposing the index per pixel, reads the tile size once per ROI, and uses size_t tile indices and StatsInt pixel coordinates, the type Pixel2 stores.
6. buildRegularContour_nontriv allocates borderImage where it called paddedImage.allocate() a second time (contour.cpp). allocate() writes from the file's current position, which set_at() leaves just after the last pixel written, so every slot after that pixel was zeroed before the trace. Pixels streamed in raster order lose nothing; pixels streamed tile by tile lose those of an earlier tile that lie lower in the box, so a ROI spanning tiles got a contour, PERIMETER and EDGE_* values that depended on the tiling.
7. A test, below.

Tests

test_2d_ooc_2d_tiled_matches_untiled_invariant (tests/python/test_2d_ooc_invariant.py) featurizes the whole 2D feature surface out-of-core on a tiled TIFF and on a strip copy of the same arrays, and requires every column of both ROIs to agree. Both sides run out-of-core, so the comparison isolates the tile scan from the out-of-core vs in-RAM differences in #509.

• The image is 160 x 96 at 32 x 32 tiles, 5 across and 3 down, so the two tile counts differ.

• Label 1 lies inside the single tile at grid row 0, column 3. With change 1 reverted the run stops with the fix(roi buffers): refuse a pixel outside the box that sizes its buffer #483 "lies outside the bounding box" error.

• Label 2 spans tiles: a column reaching down to y 60 and an arm to its right ending at y 50, so the arm's last pixel is the last one streamed. With change 6 reverted, PERIMETER, EDGE_*, CIRCULARITY and the enclosing-circle diameter of label 2 diverge (PERIMETER 259 vs 278).

• With change 2 reverted the test fails too: since change 5 places pixels by the loop's row and col, the old bounds and stride put label 1 outside its box on the 5 x 3 grid.

Change 4 has no reachable failure to test.

The full gtest suite passes, pytest tests/python/ passes apart from the 7 Arrow tests a build without Arrow cannot run, and an ASan+UBSan run of both is clean.

[vjaganat90]

LGTM. This settles everything I raised on #487:

• The contour line is fixed, with its defect comment.
• The load_tile commit no longer claims a fix.
• The redundant close() is gone.
• The new invariant compares a tiled file with an untiled copy on a grid that isn't square.

I built the head and ran mutations against test_2d_ooc_2d_tiled_matches_untiled_invariant. It fails on each of them:

• Column taken from the row (the original bug): label 1 stops with the #483 bounding-box error.
• paddedImage.allocate restored at contour.cpp:710: label 2 diverges. PERIMETER is 259 vs 278, and EDGE_*, CIRCULARITY, FRACT_DIM_PERIMETER and the enclosing-circle diameter differ too.
• The old transposed bounds and stride.
• The wrong stride alone.

test_2d_ooc_invariant.py passes, and so does the full pytest tests/python/, apart from the Arrow tests my build doesn't include.

Nits

• The description says change 2 can't be pinned by a test. Since 5fcd5b5 places pixels by the loop's row and col, it can: on the 5 × 3 grid the old bounds and stride put label 1 outside its box, and this test fails. Please drop that sentence.
• 3299303 is an empty ci: commit. Main lands PRs by rebase, so it would reach main as is. Please drop it before merge.

[darkclad]

Fixed The description no longer says change 2 can't be pinned; it now states the 5 × 3 grid test fails with change 2 reverted. The empty ci: commit is dropped, so the branch ends at e2895e9.