[cdmaestas]

Summary

Two commits:

Hash files up to two windows whole in partial_hash

• For 64 KiB < size ≤ 128 KiB the partial hash read only the first 64 KiB, so same-sized files differing past that collided and cdm dupes (without --verify) reported false duplicates. Such files are now hashed whole; bytes_read() matches.
• Domain tag kept at cdm-partial-v1: every other size hashes identically. Bumping it would have left reused v1 rows (scan reuses on unchanged size/mtime) incomparable with new ones, silently hiding duplicates.
• Schema 3 migration clears only the partial hashes in the affected size range, so the next --checksum scan recomputes them. Read-only connections (MCP) will ask for cdm doctor once to upgrade.

Support nested roots, report scan throughput, never hash the index itself

• Overlapping roots (~ and ~/src): one row per file, owned by the most specific root; hashes shared; forget hands rows to the enclosing root.
• Scan summary and progress report entries/s and hash throughput; --progress [SECS] for non-tty logs.
• Index, -wal and -shm are recorded but never opened for hashing (closing a descriptor dropped SQLite's POSIX locks → SIGBUS in walFindFrame).

Test plan

• pytest -q --cov=cdm --cov-fail-under=80: 191 passed, 1 skipped (mcp extra not installed)
• ruff check . clean
• First commit alone: 175 passed
• New tests/test_hashing.py: 100 KiB files differing at byte 80,000 get different partial hashes (collided before the fix); bytes_read matches actual reads; migration clears only affected rows

🤖 Generated with Claude Code