feat: Strategic Constraint Nodes for Golden Path directional control

Context

The Golden Path synthesis in DreamService.synthesizeGoldenPath() ranks OPEN issues by a hybrid priority score:

priority = (1/(semantic_distance + 0.1) × SEMANTIC_WEIGHT) + (structural_edge_weight × STRUCTURAL_WEIGHT)

This is purely reactive — the frontier embedding is computed from the 2 most recent session summaries, so the swarm follows its own inertia. The only mechanism for intentional steering is mutate_frontier, which adjusts edge weights but cannot express strategic vetoes or forced inclusions.

Problem

Three directional control patterns are currently impossible:

1. No-Fly Zones: "Don't touch Grid multi-body until the selection model is stabilized" — there's no way to exclude an entire component area from Golden Path candidacy without manually closing all related tickets.
2. Priority Overrides: "The Client.com migration is the #1 business priority regardless of graph topology" — strategic business priorities can't override mathematical rankings.
3. Explore Zones: "Stress-test the Canvas Worker integration even though no tickets exist" — the Golden Path is purely convergent (optimizes toward existing high-score nodes). There's no mechanism for intentional divergent exploration.

A2A Context (Fat Ticket)

Proposed Design

New node type STRATEGIC_CONSTRAINT with the following schema:

{
    id: 'constraint:no-fly-grid-multibody',
    type: 'STRATEGIC_CONSTRAINT',
    name: 'Grid Multi-Body Stability Hold',
    properties: {
        constraint_type: 'NO_FLY_ZONE',  // or 'PRIORITY_OVERRIDE' or 'EXPLORE_ZONE'
        target_pattern: 'grid.*multi.*body',  // regex matched against node IDs/names
        reason: 'Selection model regression risk — hold until #9850 is resolved',
        created_by: 'tobiu',
        expires_at: '2026-05-01T00:00:00Z',  // optional TTL
        active: true
    }
}

Integration Points

1. synthesizeGoldenPath() — After the hybrid SQL query returns scoredNodes, apply constraint filtering:

   • NO_FLY_ZONE: Remove matching nodes from candidacy entirely
   • PRIORITY_OVERRIDE: Inject matching nodes at the top with maximum score
   • EXPLORE_ZONE: Query the graph for nodes matching the pattern even if they're not OPEN issues, and add them as exploration candidates

2. mutate_frontier MCP tool — Extend to accept constraint_type as an optional parameter. When provided, creates a STRATEGIC_CONSTRAINT node instead of adjusting edge weights.

3. sandman_handoff.md — Add an "Active Strategic Constraints" section listing all non-expired constraints so agents understand the boundaries.

Avoided Pitfalls

• Constraints must have optional TTL (expires_at) to prevent permanent strategic drift. The runGarbageCollection() method should prune expired constraints.
• NO_FLY_ZONE must not delete or modify the underlying nodes — it only filters them from Golden Path candidacy. The data remains for future analysis.
• PRIORITY_OVERRIDE should be used sparingly — it bypasses the mathematical ranking entirely.

Verification Plan

1. Unit test: Create a STRATEGIC_CONSTRAINT node with NO_FLY_ZONE, run synthesizeGoldenPath(), assert the matching nodes are excluded from the top-3 output.
2. Unit test: Create a PRIORITY_OVERRIDE constraint, assert the target node appears at position 1 regardless of its mathematical score.
3. Integration: Verify constraints surface in sandman_handoff.md under a dedicated section.