fix: guarantee per-term BFS seed diversity in query (fixes #1445)

[nokternol]

Summary

Fixes #1445. _pick_seeds' gap_ratio cutoff discards any BFS-seed candidate
scoring below 20% of the top score. On a multi-term natural-language query,
one term's incidental exact label match on a node that is otherwise
unrelated to the query's intent (e.g. a common word that also happens to be
used as a field name or identifier elsewhere in the corpus) scores ~1000x
higher than any substring match on the query's other, actually-relevant
terms (_EXACT_MATCH_BONUS vs _SUBSTRING_MATCH_BONUS). The cutoff then
silently discards every one of those substring-tier candidates as seeds, so
the traversal only ever explores the neighborhood of the one unrelated exact
match — and query returns confidently-wrong results with no signal that
anything went wrong.

This is the same shape as #1445's reproduction: a vague query that doesn't
name a target symbol seeds from unrelated "concept-dense" nodes instead, even
though the target node is present in the graph and would have matched on a
weaker (substring) tier.

Fix

_pick_seeds now optionally accepts the graph and the tokenized query terms.
When supplied, it guarantees at least one seed per distinct query term that
has any match at all, so one term's incidental collision cannot starve out
the others. Ties within a term are broken by node degree (structural
centrality), so an isolated incidental match doesn't out-rank a real,
well-connected hub for that term.

Both new parameters default to None; existing callers that don't pass them
see byte-identical behavior — verified by a new test
(test_pick_seeds_without_diversity_args_is_unchanged) alongside the existing
#897 test suite, which is unmodified and still passing.

Testing

• Added test_pick_seeds_diversity_recovers_starved_term, reproducing the
  exact failure shape from BFS path drift on vague Chinese queries: 40-node corpus, 12 unrelated nodes returned for "rate_limit_check 怎么用？" #1445 with a minimal graph (one node that's an
  exact match for an unrelated term, one node that's a substring match for
  the actually-relevant term via an edge from a third node) — confirms the
  target is discarded without the fix and recovered with it.
• Full suite: 74 passed (uv run --group dev pytest tests/test_serve.py -q).
• ruff check clean on both changed files.

Scope note

While investigating this, I found (and locally patched, but am not
including here) two related but separable issues on the same code path,
happy to open as follow-ups if useful:

1. A hub-degree threshold in _bfs (max(50, p99_degree)) that, once seed
   diversity is fixed and BFS starts from more seeds, lets shared
   infrastructure nodes below that floor flood the output. The right general
   fix isn't obvious (a repo-specific floor is a blunt hack), so I left it out
   of this PR rather than guess at the right heuristic.
2. A keywords-style node metadata field so enrichment can attach
   paraphrase/alias surface beyond a bare identifier label — this is a real
   feature addition (plus a matching change to the trigram prefilter in
   _node_search_text, which silently excludes any field not indexed there),
   not a bug fix, so it felt out of scope for a PR against this issue
   specifically.

Happy to adjust the approach if you'd prefer a different fix shape (e.g.
tuning gap_ratio per-term instead of a hard per-term guarantee, or a cap on
total diversity seeds for very long queries).

[mark-butterworth-lookinglocal]

Sorry, I'll get Claude to redo this PR tonight... I'm not sure how the patch got so messed up.

[safishamsi]

Merged into v8 as d56ee83 (your authorship). Verified with a #1445-shaped graph: a multi-term query that used to collapse to a single incidental leaf (['home']) now seeds the whole relevant cluster (['home','alert','dash','grid','kpi']). Backward-compatible (G/terms are optional kwargs), and the degree tie-break is a good call so an isolated incidental match doesn't out-rank a real hub. Also partially addresses the seed-hijack in #1602 (this is the seed-selection half; the scoring/term-coverage half is still open there). Full suite 2855. Ships next release — thanks.