test(agent): pin Windows HOME variables in tilde-dependent tests

[LceAn]

What & why

On POSIX, expanduser("~") honors HOME; on Windows it prefers USERPROFILE, then HOMEDRIVE/HOMEPATH. Tests that set only HOME can therefore resolve ~ against the real Windows profile instead of their temporary directory.

Change

Apply the portable test-home setup to all seven reviewed tilde-dependent tests:

• set HOME and USERPROFILE to the temporary test directory;
• clear HOMEDRIVE and HOMEPATH fallbacks;
• compare the two mixed-separator path results semantically with Path.

No production code changes are involved.

Validation

The seven focused behaviors pass on native Windows (six pytest targets plus direct MCP certificate-path verification); lint and diff checks pass.

Related / duplicate analysis

Complementary to #67196, which changes canonical runner/Hindsight isolation. Not a duplicate of #34374/#61571/#61668: those change production image-routing recognition, while this PR only repairs test fixtures.

[teknium1]

Thanks for the focused Windows test-isolation fix. The two target failures are real on current main: tests/agent/lsp/test_workspace.py:136-139 sets only HOME before agent/lsp/workspace.py:41 calls os.path.expanduser, and tests/agent/test_context_references.py:334-355 similarly exercises the home-sensitive guard at agent/context_references.py:366.

Problems

• The same test defect remains in unmodified sibling paths: tests/agent/test_shell_hooks_consent.py:203 → agent/shell_hooks.py:880; tests/agent/test_image_routing.py:541 → agent/image_routing.py:122; tests/cli/test_cli_file_drop.py:214 and tests/cli/test_cli_image_command.py:76 → cli.py:2850; tests/tools/test_mcp_client_cert.py:87 → tools/mcp_tool.py:992. Each relies on HOME alone while asserting ~ resolves under its temp directory.

Suggested changes

• Extend this portable-home setup to those sibling tilde tests. tests/tools/test_checkpoint_manager.py:57-60 is existing repository precedent.

Automated hermes-sweeper review.

[alt-glitch]

This was generated by AI during triage.

Related to #49118: this PR now expands the same Windows HOME-profile isolation to seven tests, so it is a broader, complementary test fix rather than a duplicate.

[LceAn]

Updated per review. Added the portable HOME/USERPROFILE setup to all five named sibling tests (seven total), and changed the two mixed-separator string assertions to Path semantic comparisons.

Focused validation: all 7 target nodes passed.

The broader canonical seven-file run still exposes independent existing Windows clean-env baselines: CLI files fail collection because the runner removes USERPROFILE/LOCALAPPDATA before cli import, test_image_routing.py has six unrelated absolute-Windows-path parser failures, and the runner itself raises GBK UnicodeEncodeError while printing ✓. I kept those separate from this focused HOME/tilde PR.

[LceAn]

Maintenance update: rebased onto current main, normalized all seven touched files back to LF, and refreshed the body with the related-PR analysis. The seven reviewed portable-home fixtures remain intact; the diff is now +29/-5. #67196 is complementary, and the duplicate image-routing PRs target production regex code rather than these fixtures.

[LceAn]

Second-pass maintenance: kept the seven per-test HOME/USERPROFILE fixtures as the narrow fix; no global fixture was introduced. The broader runner and production image-routing PRs remain complementary, not duplicates.