fix: name community splits from their members and dedupe community names

[SHudici]

Problem

Two related defects make community output hard to trust on real graphs:

1. Splits are opaque. _split_oversized names every shard "<parent>-sub<N>" and hardcodes "cohesion": 0.0. On a mid-size production graph (~7k nodes, 103 communities), 26 of the 103 communities were indistinguishable services-load-subN entries, all reporting zero cohesion — the numbers looked like detection failures, not results.
2. Names collide. Nothing enforces uniqueness: the same graph carried five collision groups (services-normalize ×4, services-job ×3, services-fake ×3, services-sql ×2, services-testclientlazyinit ×2 — 14 communities sharing 5 names). get_community resolves by name match, so 9 of those 14 were unreachable by name.

Fix

• _split_oversized names each shard from its own members via the existing _generate_community_name (fallback "<parent>-<id>" only when members yield nothing), and computes real cohesion for all shards in one _compute_cohesion_batch pass over the full edge set — same cost profile as the top-level detectors.
• detect_communities runs a new _dedupe_community_names pass after splitting: within a collision group the largest community keeps the bare name; every other member is suffixed with its most distinctive keyword (skipping keywords already in the name and candidates that would collide with any existing name), with a deterministic numeric fallback.

Measured effect (same production repo)

Re-ran community detection on the same graph (103 communities before and after — the partition is untouched, only naming and cohesion change):

metric	before	after
-subN placeholder names	26	0
communities reporting cohesion 0.0	26	0
duplicate-name groups	5 (14 communities)	0
communities unreachable via get_community by name	9	0

The 26 former services-load-subN shards now carry member-derived names with real cohesion, e.g. services-upload (85 members, 0.29), services-rows (74, 0.42), lx-raw (35, 0.27), iag-market (8, 0.33), afklm-parse (7, 0.41).

Testing

New tests: a dumbbell-graph oversized community splits into member-named shards (no -subN), each with exact expected cohesion (15/16) and parent lineage preserved; dedup keeps the bare name on the largest community, suffixes the rest by keyword, skips keywords already in the name, avoids colliding with existing community names, falls back to a numeric suffix, and leaves unique names untouched. Full suite passes.

Composes with #600 (community naming quality) but does not require it; cut independently from main.

🤖 Generated with Claude Code