repomap: boost co-changed files in personalization via git history

[De-Cri]

Closes #5203.

When editing group.py in brian2, the repomap correctly surfaces its dependencies (codeobject.py, variables.py, base.py). It does not surface neurongroup.py or synapses.py, even though both import from group.py and are almost certainly affected by any change to it. The reference graph edges go neurongroup → group and synapses → group. PageRank distributes weight along group.py's outgoing edges, so the subclasses — which point toward it — never surface.

In 5 years of brian2 history, group.py has been committed alongside synapses.py 111 times and neurongroup.py 76 times. Git history sees the coupling that the reference graph misses.

Implementation

New private method _git_cochange_partners(abs_fname):

• Runs git log --no-merges -n500 --name-only --format=COMMIT -- <file>
• Counts file co-occurrences per commit
• Returns {rel_fname: rate * log(1 + count)} for pairs above min_count=2
• Returns {} gracefully if git is unavailable, repo has no history, or command times out (5s)

In get_ranked_tags(), after the existing personalization loop:

all_rel_fnames = {self.get_rel_fname(f) for f in fnames}
for fname in chat_fnames:
    for partner_rel, score in self._git_cochange_partners(fname).items():
        if partner_rel in all_rel_fnames:
            personalization[partner_rel] = (
                personalization.get(partner_rel, 0) + score * personalize
            )

No new dependencies. No new flags. No API change. Purely additive to the existing ranking.

Benchmark

Ground truth: held-out recent commits. For each commit that modified 2+ source files, given one file — does the tool surface the others?

	Recall@5
Without this change	24%
With this change	88%

(brian2tools, 25 held-out commits, 233 training commits.)

The signal degrades gracefully on repos with little history — below ~50 commits there are few reliable pairs and the boost is near zero.

[CLAassistant]

All committers have signed the CLA.