[wenzeslaus]

The temporal tests needs separate environment like all the other tests, but env can't be explicitly passes as for the rest of the API, so this uses pytest's monkeypatch to work around that.

Details

Problem

The fixture added in #7714 creates its sessions with gs.setup.init(...) without env, which sets the session up in the global os.environ. init() calls setup_runtime_env() unconditionally, which prepends the GRASS executable paths to PATH and never removes them again, so every such call grows PATH permanently for the rest of the pytest process (measured: +404 characters per call). pytest runs everything in one process, so later tests which copy os.environ inherit it.

Fix

Create the sessions with env=os.environ.copy(). tgis reads the session from os.environ and cannot take an env, so the second session is mirrored there with monkeypatch, which restores the environment when the module finishes. test_project_without_temporal_database in the same file already does this, so this only makes the fixture consistent with it.

AGENTS.md gets the rule and the tgis exception documented in its Writing Tests section.

What it broke

r.coin has been failing in the OSGeo4W (Windows) CI on main since #7714 was merged, with nothing in r.coin changed. It is not flaky: every run whose base contains c7fc4ea fails, every run without it passes (20 runs checked). r.coin runs r.stats in a subprocess rather than computing the table itself, and on the inherited PATH cmd.exe could no longer resolve it:

'r.stats' is not recognized as an internal or external command,

so r.coin printed an empty table and the test saw [(0, 1, 0), (0, 2, 0), (0, 3, 0)]. It is the only tested tool which runs another tool, which is why it alone is affected.

Follow-ups, not in this PR

• The ultimate fix is for grass.temporal to accept env; the mirroring here could then go away.
• gs.setup.init() re-prepending to PATH on every call is worth fixing on its own, since it is what turns one env-less init() into a process-wide leak. 18 other env-less call sites remain.

Verification

Fixture before/after in an identical harness: PATH growth goes +444 to +0. The mirroring pattern was verified against a local build: tgis.init() and tlist_grouped() still work through the mirrored environment and os.environ is restored. The Windows threshold at which cmd.exe stops resolving r.stats was not reproduced locally (no Windows machine, and PATH is not printed in CI logs); CI on this PR is the check for that.

The investigation and the patch were done with the assistance of Claude Code.

[wenzeslaus]

Issues opened for the individual causes found in this investigation:

• [Bug] grass.script.setup: Repeated init() on the same environment grows PATH, PYTHONPATH, and the library path variable #7738 — repeated gs.setup.init() on the same environment grows PATH, PYTHONPATH, and the library path variable (the accumulation this PR works around)
• [Feat] grass.temporal: Allow passing env instead of using os.environ #7737 — grass.temporal should accept env (would remove the need for the mirroring in this PR)
• [Feat] lib/gis: Resolve spawned programs with find_program instead of cmd.exe on Windows #7736 — lib/gis spawns through cmd.exe on Windows, which is what made the overlong PATH fatal for r.coin
• [Feat] lib/gis: Return the child exit status from G_popen_close #7735 — G_popen_close() discards the child exit status, so r.coin could not notice r.stats failing
• [Bug] r.coin: Prints a category 0 table and exits with success when r.stats produces no output #7734 — r.coin prints a category 0 table and exits with success when r.stats produces no output

Any of these would help to avoid the issue this PR solves, but each makes sense separately, but I can't create PRs for all at this point.