Implement InteractionGraph with decay, GC, and neighbor queries

[rebelmachina]

Epic: #64 (AffinityEngine)
Depends on: #65 (crate scaffold)

What

Implement the InteractionGraph data structure in crates/arcane-affinity/src/interaction_graph.rs:

• EntityPair — canonical ordered pair (min(a,b), max(a,b)) avoiding duplicate entries
• InteractionGraph struct with HashMap<EntityPair, f64> storage
• Methods: new(), record_interaction(), tick() (decay + GC), get_weight(), neighbors(), remove_entity(), pair_count()
• neighbors() is the hot path — called once per entity per clustering tick

Key design decisions

• Start without secondary index for neighbors(). Add HashMap<Uuid, Vec<Uuid>> only if profiling shows it's a bottleneck.
• GC runs every gc_interval ticks, removing entries below gc_threshold
• Memory estimate: 6,000 entities with 10% interacting ≈ 30k-50k pairs ≈ 1.2 MB

Acceptance criteria

• All methods implemented per spec
• Unit tests for: record/add-to-existing, canonical ordering, decay, GC, neighbors, remove_entity, pair_count

Spec

See arcane-engine repo: in_08_01_interaction_graph.md

[rebelmachina]

Implemented in PR #77. InteractionGraph in crates/arcane-affinity/src/interaction_graph.rs: canonical EntityPair ordering (min, max), additive record_interaction, exponential decay + GC via tick(decay_factor, gc_threshold, gc_interval), neighbors() iterator, remove_entity(). 7 unit tests including self-interaction guard and exact GC threshold.