Stop dft_rast_classify() modifying the caller's raster (#89)

[NewGraphEnvironment]

Summary

• dft_rast_classify() no longer modifies the caller's raster. It called the in-place terra::set.cats() on the object it was given. Now coltab<- runs first, which deep-copies, and the levels are set on that copy. The output is identical() to 0.19.0 (cats() and coltab()), and no extra copy is made.
• A new test covers file-backed, in-memory and list-element inputs. It goes red on the old order with 6 failures. The workaround comment in test-dft_accuracy_sample.R is removed.
• Release v0.19.1: NEWS, and a CLAUDE.md note on measuring peak RSS for short calls.

Related Issues

• Fixes dft_rast_classify() mutates the caller's raster in place (set.cats on the input) #89
• Filed on the way: dft_rast_classify() returns empty levels when its input is already a factor #91 (a raster that is already a factor gets empty levels; pre-existing, not fixed here)
• Relates to NewGraphEnvironment/sred-2025-2026#16

Scale check (BULK, per CLAUDE.md)

BULK classified_2017.tif (169,248,352 cells), terra 1.9.50. Kernel max RSS from /usr/bin/time -l, two reps each:

variant	input	max RSS (GiB)	caller untouched
0.19.0	in memory	4.21	no
this PR	in memory	4.21	yes
deepcopy() first (the issue's proposal)	in memory	5.46	yes
0.19.0 / this PR	file-backed	1.05 / 1.05	no / yes

deepcopy() would add 1.25 GiB per year classified in memory, so it was not used. The classify call itself takes 0.8–0.9 s on both main and this branch. The first attempt used a 2 s ps sampler, which cannot catch a sub-second peak: it read 0.26 GiB. I first blamed sampling the wrong PID, and retracted that after checking. The archive README records the full chain.

Test plan

• devtools::test(): [ FAIL 0 | WARN 0 | SKIP 15 | PASS 1318 ]
• New test red on the old order (6 failures), green on the fix
• /code-check: three rounds, all clean. Reviewers confirmed that coltab<- copies at the terra 1.8-10 floor, that old and new orders give identical output on nine input shapes, and that no caller relied on the mutation.
• Changed lines lint clean

🤖 Generated with Claude Code

https://claude.ai/code/session_01PRhUJsuKABLfpBGktPoiBN