feat: Re-implement STAN Stigmergic A* Navigation (Resolves #258 feedback)

[web3guru888]

Fixes previously reverted Feature: Implement STAN (Stigmergic A* Navigation) into Knowledge Graph

This PR successfully re-implements the Stigmergic A* Navigation capabilities originally proposed in #258, but definitively resolves all critical flaws pointed out by maintainers via a direct port mapping from the tested architecture of the STAN research paper by VBRL.ai.

What's Fixed:

🔴 1. Schema Migration No Longer Breaks Existing DBs:
A safe _apply_migrations method was incorporated which conditionally handles ALTER TABLE triples ADD COLUMN pheromone_level REAL DEFAULT 0.0. It correctly ignores the operation if older nodes are already populated with schemas locally.

🔴 2. Entity ID Double-Normalization Erased:
In STAN's modified A*, we query using get_neighbors_with_pheromones, returning correctly formatted semantic ids. We no longer force double conversions for path output/input looping calls.

🔴 3. Evaporation (Decay) Routines Built In:
Unlike the initial proposal which only deposited pheromones continuously leading to full network "lock-in," STAN actively includes 10% evaporation (rho=0.1). I added evaporate_pheromones(decay_rate=0.1) dropping paths uniformly, keeping learning strictly elastic.

🔴 4. Pathfinding O(n^2) Scaling Fixed:
Instead of repeatedly deep copying path matrices through $A^*$ states, segments are captured as raw (from, to, predicate, direction, cost, tau) dict instances inside heapq.

🔴 5. Active Execution Engine (MCP Handlers):
This code is fully active, with two new endpoints injected directly into your mcp_server.py:

• mempalace_kg_stigmergic_astar: Retrieves discounted route patterns.
• mempalace_kg_deposit_pheromone: Registers pheromones with an underlying modular tick constraint (evaporate operations trigger unconditionally at % 50 == 0 internal MCP calls).

Local Testing Completed 🟢:

Using rigorous unit tests natively extending MemPalace dependencies, we successfully mapped that route selections switch to longer graph distances accurately when underlying stigmergic weights reduce the calculation below base defaults without edge leaks.

(Replaces reverted Pull Request #259).

[bensig]

Appreciate the thoroughness here — addressing all the feedback from #259 shows real effort. The stigmergic pathfinding concept is interesting.

However we're focused on stabilizing the core right now (search, mining, MCP reliability) and this adds significant complexity to the KG with new schema migrations, MCP endpoints, and a novel algorithm. We'd rather get the foundation solid before adding research-grade features.

Would you consider publishing this as a standalone extension or plugin? That way people can opt in without it being in the core package. And we'd value your help reviewing PRs — you clearly understand graph algorithms well.