Verify Tiers 2/3 of the SPAWNED inference rule against real concurrent-subagent Claude Code sessions

[antejavor]

Context

The three-tier temporal-inference rule for (:Action)-[:SPAWNED]->(:Agent) (Map: Harness-grounded graph model across the Context Graph family, ticket #277) has three tiers:

1. Filter open, agent-spawning tool calls.
2. If multiple are open, disambiguate by matching agent_type against tool_input.subagent_type.
3. Genuine ambiguity (same type, concurrent) resolves to an honest null, never a guess.

Surfaced while scoping a new map for live-harness verification of this rule. Tier 1 (a single subagent spawn) was scoped in as that map's destination. Tiers 2 and 3 need a real Claude Code session where the model spawns multiple subagents in overlapping time — reliably forcing that shape via prompting alone is a harder, less deterministic problem than Tier 1, and risked stalling the whole effort on prompt-engineering flakiness rather than proving the core mechanism works at all. Ruled out of scope for that map; tracked here as a deliberate follow-up once Tier 1 verification is solid.

Scope

• Design a prompt/scenario that reliably gets a real Claude Code session to spawn two subagents with overlapping lifetimes (same and different agent_types, to exercise both tiers).
• Extend whatever Tier 1 verification tooling lands from map Map: Harness-grounded graph model across the Context Graph family #275's follow-on to also assert Tier 2/3 behavior.