fix: NaN confidence filter silently rejected all graph_accel traversals

[aaronsb]

Summary

• Root cause: graph_facade.py passed float('nan') as the min_confidence parameter to graph_accel SQL functions when no confidence filter was requested. In IEEE 754, NaN >= x is always false, so the Rust confidence filter rejected every edge with a loaded confidence value — silently returning empty results from BFS neighborhood, shortest path, and subgraph extraction.
• Fix: Pass None (SQL NULL) instead, which maps to Rust's Option::None and bypasses the confidence filter entirely.
• Impact: All three accelerated traversal functions (graph_accel_neighborhood, graph_accel_path, graph_accel_subgraph) were affected. Without this fix, any graph where edges have confidence properties (the common case for semantic edges from ingestion) returned zero results.

How it was found

Searching "Negentropic Viability Model" in the web workstation returned only 1 concept with 1 adjacent at 4 hops deep. Investigation initially focused on the GUC edge-type filtering from PR #339, but the actual bug predates that PR — it was present from the original graph_accel facade integration. The GUC filtering fix (correctly reducing edge count from 19K to 3.9K) just made the empty results more noticeable since there was less data to mask the issue.

The IEEE 754 NaN trap

# Before (broken): NaN as "no filter" sentinel
conf_param = min_confidence if min_confidence is not None else float('nan')
# → Rust: Some(NaN) → edge.confidence >= NaN → always false

# After (fixed): NULL = no filter
# → Rust: None → skip filter entirely

Test plan

• MCP concept related returns 83 concepts at depth 2 (was 0)
• MCP concept related at depth 4 returns extensive results (81K+ chars, was 0)
• Direct SQL comparison: NULL returns 83 results, NaN returns 0
• Rust core two-phase loading tests pass (cargo test)
• Web workstation: search "Negentropic Viability Model" shows rich neighborhood at 4 hops