Summary

• Make norm_dir_prefix (and its copy-pasted twin in system.py:remove_memory_dir) append os.sep instead of a hardcoded "/", so the trailing-separator prefix matches norm_path's native-separator output on Windows.
• Fold the inline prefix.endswith("/") block in system.py into a single norm_dir_prefix(resolved) call so the rule lives in one place.

Why

norm_dir_prefix produces a str.startswith prefix from a path that has been through Path.resolve() — which returns native-separator form (C:\Users\foo on Windows, /Users/foo on POSIX). The hardcoded "/" append yielded a mixed-form prefix C:\Users\foo/ on Windows that no chunk source path could ever match under target.startswith(prefix).

Concretely, on Windows this broke:

• resolve_owning_memory_dir — every chunk classified as orphan (kind=None).
• memory_dir_stats — per-dir bucketing returned 0 chunks for every dir.
• /api/sources route — sources list missed kind attribution.
• /api/memory-dirs/remove with delete_chunks=true — removed 0 rows.

POSIX was unaffected because os.sep == "/" there; the bug only showed up once the Windows leg of the CI matrix went informational (#634 / PR #638).

Approach

Symmetric, minimal:

# engine.py:norm_dir_prefix
base = norm_path(p)
if not base.endswith(os.sep):
    base += os.sep
return base

norm_path itself is not touched — both the write side (sqlite_backend.upsert_chunks) and the read side (startswith comparisons) round-trip through the same Path.resolve() + Unicode NFC normalisation, so drive-letter case, 8.3 short names, and the \\?\ long-path prefix all collapse to a single canonical form via the OS. No DB migration is needed; the persisted chunks.source_file column is bit-for-bit identical to today on every platform.

This mirrors the bug class fixed by #716 / 82feed1 (fix(config): support Windows backslash paths in categorize_memory_dir), but the analog here is os.sep — not .replace("\\", "/") — because the surrounding comparison surface in engine.py and web/routes/* is already native-form on both sides.

What's not in this PR

• No norm_path change. Forcing forward-slash inside norm_path would invalidate every existing chunks.source_file row on a populated Windows install (Windows CI is currently informational, but the contract is shared with 16 call sites — not worth changing for marginal benefit).
• No new tests. The 23 currently-failing Windows tests in test_indexing_engine.py::TestResolveOwningMemoryDir (5), ::TestMemoryDirStats (18), and test_web_routes.py::TestSources / ::TestRemoveMemoryDirChunkCleanup / ::TestUploadsUsage are the validation surface — they construct paths via tmp_path (native form on every OS) and assert via Path equality, so they go green on Windows-CI once the prefix uses os.sep. POSIX stays bytewise green.
• No test-side path-portability sweep. A handful of substring assertions in test_wiki_cmd_override.py still hardcode /skills/... literals — covered separately as a sibling of test: normalize path separators in 7 cross-platform assertions (#643) #711.

Test plan

• uv run pytest packages/memtomem/tests/test_indexing_engine.py packages/memtomem/tests/test_web_routes.py -m "not ollama" — 253 passed (POSIX no-op confirmed).
• uv run ruff check packages/memtomem/src/memtomem/indexing/engine.py packages/memtomem/src/memtomem/web/routes/system.py — clean.
• uv run ruff format --check on the two changed files — clean.
• grep -rn 'endswith("/")\|prefix.*+= *"/"' packages/memtomem/src/ --include="*.py" — zero hits remaining.
• Windows-CI leg: 23 listed failures (TestResolveOwningMemoryDir x5, TestMemoryDirStats x18, TestSources/TestRemoveMemoryDirChunkCleanup/TestUploadsUsage) turn green.

Closes #647. Refs #634 (Windows + macOS CI matrix umbrella).

🤖 Generated with Claude Code