fix: prevent coverage data corruption from artifact name collisions - #1116

I hope this fixes the failure we've seen occasionally in CI after multiprocessing in the test suite.

🤖 AI text below 🤖

Problem

The coverage (combine) job intermittently fails with:

nox > coverage combine
Couldn't use data file '/home/runner/work/nox/nox/.coverage': database disk image is malformed
nox > Command coverage combine failed with exit code 1

(e.g. this run).

Cause

macos-latest (arm64) and macos-15-intel (x86_64) both run Python 3.12, and the coverage filename is built only from sys.platform + version — so both jobs produce .coverage.pypi.darwin.3.12. The cover job then downloads artifacts with merge-multiple: true, which extracts them concurrently into one directory. Two jobs writing the same path can tear the SQLite data file (→ "database disk image is malformed"), or one silently clobbers the other's coverage. The timing dependence is why it only fails occasionally; the parallel/xdist tests just made the corruption more visible, they aren't the trigger.

Fix

• Add platform.machine() to the tests/conda coverage filenames so arm64 and x86_64 no longer collide (also restores the mac coverage that was being silently dropped).
• Drop merge-multiple: true so each artifact lands in its own coverage-* subdirectory; the cover session now globs those subdirs and passes them to coverage combine (with a ["."] fallback for local runs).

No change to the fast test setup — xdist, the sysmon core, and run.patch=["subprocess"] are untouched. The 100% --fail-under gate is unchanged.

Verified locally that coverage combine <dir> <dir> reads data files inside per-artifact subdirs and combines cleanly.