Georef output contract: alpha band for grayscale, srcnodata as mask fallback (#56)

[NewGraphEnvironment]

Summary

• Breaking: every fly_georef() output now gets -dstalpha. Grayscale goes from 1 band with NoData = 0 to 2 bands, Gray + Alpha, no NoData. RGB stays 4.
• mask = "border" together with srcnodata is now accepted. srcnodata becomes the fallback for frames whose mask declined. fly_georef_warp_opts() was always an else if, so the old refusal guarded a pair GDAL never received.
• georef_one() now warns when a mask declines and no fallback was given, because that frame's collar reaches the output as image data.

Related Issues

• Closes Georef output contract: a real alpha band for grayscale, and a supported mask-or-nodata fallback #56 (which absorbed Masking adds a band on Windows for grayscale, so "output band counts do not change" is platform-conditional #68 and No supported way to ask for "mask the collar, fall back to nodata" #69)
• Downstream: Grayscale georef outputs become Gray + Alpha: COG writer drops the alpha colorinterp stac_airphoto_bc#36. Its COG writer refuses gray+alpha until it creates the in-memory GTiff with alpha="YES", and grayscale frames then need republishing.

What was measured

Measured on the 182 grayscale calibration thumbnails, on sf's GDAL 3.8.5. Producer: data-raw/mask_measure-interior_zeros.R.

old -dstnodata 0 path	frames where true black was rewritten as 1	max share of a frame
axis-aligned warp	161 of 182	3.6%
30° warp	41 of 182	0.26%

• No pixels were deleted on the default path.
• The 0→1 shift from Masking adds a band on Windows for grayscale, so "output band counts do not change" is platform-conditional #68 is not specific to Windows. sf's GDAL 3.8.5 does it too, without printing anything.
• The mask declined 0 of 10,105 thumbnails, so on thumbnails the fallback changes nothing. Where it matters is non-8-bit scans.

Test plan

• devtools::test(): 2585 pass, 0 fail, 0 warn (macOS)
• Mutation checks. Dropping -srcnodata turns the fallback test red. Restoring -dstnodata 0 next to the alpha band turns the collision test red on its values, not just on band count.
• pkgdown::check_pkgdown() clean
• Three-platform CI. The Windows pin in test-fly_georef_mask.R is removed, and the test now asserts 2 grayscale bands on every runner. Windows is the real test of Masking adds a band on Windows for grayscale, so "output band counts do not change" is platform-conditional #68.
• Downstream COG round trip run against this branch. It failed as the table above describes, and that is filed as stac_airphoto_bc#36.

Notes

• Review: /code-check ran 3 rounds on the code commit and 3 on the docs and measurement commit. Every finding in both loops was prose or measurement scope, and all were fixed. The docs loop ended by enumerating 30 claims. Review under-counted twice: first the source-side count, then a single bearing. The archive README keeps both wrong turns.
• An issue was also filed in a private sibling that pins these warp options.
• Version bump left for /gh-pr-merge. The change is breaking and pre-1.0, so it should be a minor bump.

🤖 Generated with Claude Code

https://claude.ai/code/session_0115csb82i78r86GWFzwPzis