fix: green Windows CI — POSIX display paths after ~/, and home isolation that holds on Windows

[Brian Krabach (bkrabach)]

Why

Windows CI has failed on main for every recent run — 8 tests (e.g. run 33706023624 @ b85867c). Two independent root causes. One is a real display bug; the rest are POSIX-only assumptions in tests, one of which is also a hygiene hazard for anyone running the suite on a Windows machine.

1. routing list / show rendered mixed separators — product bug, 4 tests

_display_path abbreviated a home-relative path as f"~/{path.relative_to(Path.home())}". On Windows the remainder stringifies with backslashes, so the separator switched mid-path — in the terminal and in the JSON matrix_file field:

~/.amplifier\routing\openai.yaml
~/.amplifier\cache\amplifier-bundle-routing-matrix-test\routing\balanced.yaml

…while the CLI's own help text spells that location ~/.amplifier/routing/.... ~/ is a POSIX-style abbreviation, so the remainder now renders via as_posix() on every platform. A path outside the home directory isn't abbreviated and stays in native form — unchanged.

test_routing_shadowing.py:249, :405 · test_routing_winner_selection.py:323, :400

2. Two matrix_path assertions were POSIX-specific — test bug, 2 tests

matrix_path is a native absolute path — correctly; a user or tool may open it. .endswith("/.amplifier/routing/openai.yaml") can never hold against C:\...\.amplifier\routing\openai.yaml. Now compared as a Path.

test_routing_shadowing.py:304 · test_routing_winner_selection.py:346

3. monkeypatch.setenv("HOME", tmp_path) isolates nothing on Windows — test bug, 2 failing tests; 47 sites in 5 files with the same latent hazard

Path.home() is os.path.expanduser("~"). On Windows, ntpath.expanduser reads USERPROFILE (then HOMEDRIVE+HOMEPATH) and never consults HOME. So every SessionStore-touching test in these files wrote real records into the runner's — or a Windows developer's — actual ~/.amplifier/projects/<slug>/sessions/, and the two tests asserting not SessionStore().exists(SUB_SESSION_ID) found the record a previous checkpointing test had left there.

Fixed once: tests/conftest.py gains isolated_home (sets HOME and USERPROFILE, returns tmp_path). Each affected module opts in with a 3-line autouse fixture; its per-test HOME lines are removed so there is one mechanism, not two.

file	sites
test_timedout_session_resumable.py (the 2 failing)	12
test_session_spawner.py	22
test_resume_preserves_provider_promotion.py	8
test_resume_system_prompt.py	4
test_skills_cli.py (custom path; USERPROFILE added beside it)	1

The three non-failing files had the identical defect — their assertions merely never depended on the isolation holding.

Verification

No Windows host available, so each fix was checked against the exact values the CI log recorded, with Windows path semantics simulated via PureWindowsPath / ntpath.expanduser:

	old	new
_display_path	~/.amplifier\\routing\\openai.yaml	~/.amplifier/routing/openai.yaml (both CI values; outside-home unchanged)
matrix_path assert	.endswith(...) → False on the CI value	Path == → True (Windows and POSIX)
home isolation	HOME only → expanduser('~') = C:\Users\runneradmin	HOME+USERPROFILE → the tmp path

Linux: 1647 pass (count unchanged), deterministic and random order. Two touched test files weren't ruff format-clean on main; they were edited surgically (per-test line removals + the fixture) rather than reformatted, so the diff is the change and nothing else. ruff check clean on all 9 touched files; the 5 remaining repo-wide lint findings are pre-existing on main.

This PR should be merged only when the Windows jobs are actually green — that's the point of it.