Prune dense semantic floor so community modularity recovers

[raphasouthall]

Problem

GraphRAG community detection sat at Newman modularity Q ≈ 0.06 (coarse) / 0.05 (fine) — barely above a random partition. Communities bled into each other and the hierarchy inverted.

Diagnosis (measured, not assumed)

The blended similarity matrix is S = 0.6·semantic + 0.25·cooc + 0.15·wikilinks. A read-only parameter sweep against a copy of a ~490-note vault showed:

• Co-occurrence is not the cause. Entity document-frequency is healthy — 90% of entities (9,829/10,948) appear in a single note; the biggest hub is in 37. Co-occurrence knobs (MIN_SHARED, hub df-cap) left Q unchanged (0.067 → 0.067).
• The semantic signal is a dense floor. Off-diagonal cosine averaged 0.358, with 66% of all note pairs above 0.3. In a single-domain vault everything is mildly similar to everything, so the dominant signal connects nearly all pairs and collapses modularity.
• Thresholding the floor is the lever. Zeroing semantic edges below mean + 0.5·std lifts Q to ~0.30 coarse / ~0.28 fine with stable community counts (7 coarse / 11–12 fine across the whole sweep). top-k sparsification alone hurt (it keeps each node's top neighbours regardless of absolute weight).

config	Qc	Qf
baseline (no threshold)	0.068	0.058
co-occurrence df-cap / min-shared	0.067	0.057
mean + 0.5·std (this PR)	0.343	0.278

Change

_build_similarity_matrix zeros semantic edges below an adaptive threshold SEMANTIC_THRESHOLD_K (default 0.5, i.e. mean + 0.5·std of the off-diagonal distribution). Adaptive rather than a fixed cosine so it self-tunes to each vault's similarity spread. None disables. Guarded for n ≤ 2.

Tests

New TestSemanticThreshold: cross-cluster edges pruned while within-cluster survive, disabled mode keeps the floor, small-vault skip. Full suite: 559 passed.