Background

Indexing-related tests fail on Windows where path comparisons across different canonicalization layers (drive letter case, short-name vs long-name, separator) don't match. Symptoms include `kind=None` where `'memory'` was expected, chunks not getting deleted because their source paths don't compare equal to `memory_dirs`, etc.

Symptoms

```
FAILED test_web_routes.py::TestSources::test_kind_set_when_source_under_memory_dir
AssertionError: assert None == 'memory'

FAILED test_web_routes.py::TestRemoveMemoryDirChunkCleanup::test_delete_chunks_true_removes_matching_source_files
assert 0 == 4

FAILED test_web_routes.py::TestUploadsUsage::test_populated
assert 0 == 2
```

The pattern: chunks have a source path stored at index time; later code compares against a configured `memory_dir`. On POSIX both go through `Path.resolve()` consistently. On Windows, the resolution may differ (drive case, `\\?\\` long-path prefix, short `8.3` form) so equality fails.

Affected files (estimated)

• `packages/memtomem/tests/test_indexing_engine.py` — 11 failures
• `packages/memtomem/tests/test_web_routes.py::TestSources` / `TestRemoveMemoryDirChunkCleanup` / `TestUploadsUsage`
• Possibly: `test_dirty.py`, `test_index_debounce.py`

Fix direction

1. Audit the production code that compares `chunk.metadata.source_file` against `memory_dirs` — confirm both go through identical canonicalization (`Path.resolve()` after a consistent normalization).
2. Consider storing canonical paths at index time so later comparison is just `==`, no resolve-on-read race.
3. Test fixtures may need to use `tmp_path.resolve()` consistently to avoid `8.3` vs long-path mismatches.

Refs

• ci: add Windows runner to CI matrix #634 — Windows + macOS CI matrix expansion (umbrella)
• PR ci: matrix infrastructure + macOS gating; Windows informational (#634) #638 — matrix infrastructure (Windows currently informational)

🤖 Generated with Claude Code