feat(repo stats): dedup commits by git patch-id

[kyle-rader-msft]

Problem

lk repo stats inflates contributor counts when a repo has logically-identical commits with different SHAs — the typical case being cross-repo history migrations (rebase, cherry-pick, or graft merges), but it also shows up for any team that rebases / cherry-picks heavily.

Case study: Microsoft Agency repo

The Agency repo ingested the ESAI client code via a true git merge that preserved both parents — but a slice of the same ESAI history had also been synced in earlier as rebased copies. Git correctly attributes each SHA to its original author, but the same logical patch gets counted twice. Before this PR:

Mode	Total commits	Top contributor
lk repo stats --all	5,309	Kyle Rader: 704 (massively inflated — he drove the migration)

After:

Mode	Total commits	Top contributor
lk repo stats --all	4,471 (838 collapsed)	Hardik Kumari: 378 (correct)
lk repo stats --all --no-dedup	5,309	Kyle Rader: 704 (old behavior)

Solution

Add patch-id deduplication, on by default, with --no-dedup to opt out.

For each set of commits sharing a git patch-id --stable, keep the commit with the earliest author date (SHA breaks ties). Commits with no patch output (empty diffs, merge commits) are always treated as unique.

The summary line surfaces the count:

Total commits: 4471 (838 duplicate patches collapsed; --no-dedup to disable)

When --no-dedup is passed, the suffix is omitted.

Implementation

Refactored repo_stats() from a one-pass tally-during-git log loop into collect-then-tally. The new helpers:

• collect_raw_commits() — buffers (sha, ts, name, email) from git log.
• compute_patch_ids() — spawns git diff-tree --stdin -p | git patch-id --stable with a writer thread feeding SHAs and a reader collecting <patch-id> <sha> pairs (avoids the classic pipe-buffer deadlock).
• dedup_commits() — groups by patch-id (or own SHA for unmapped commits) and picks the earliest winner per group.

Filtering, the active-window logic, and author canonicalization all run on the deduped winner list, so existing flags (--days, --from, --top, --name, --email, --all / --first-parent) keep working as expected.

Tests

Added 5 unit tests for dedup_commits() covering: no patch-ids, basic collapse, timestamp ties (SHA tiebreak), unmapped commits remaining unique, and mixed mapped+unmapped. Full cargo test suite (38 tests) passes.

Compat

Backward-compatible feature addition — minor version bump 2.4.0 → 2.5.0 per AGENTS.md.

Existing users who want the raw SHA counts can pass --no-dedup.

Performance

Full Agency repo (~5k commits, no time filter) dedup pipeline runs in ~3s; typical time-windowed queries are sub-second. The existing ProgressMeter covers it.

Out of scope (potential follow-ups)

• A --show-duplicates diagnostic listing per-author duplicate counts
• .mailmap integration
• --since-ref to skip dedup before a given graft point