Add a wisp dispatch circuit breaker for runaway re-dispatch loops

[rockfordlhotka]

Why

Follow-up to #462 (the trim-loop runaway fix). That PR stopped the specific wedge; this one adds the guard that was missing one layer up: nothing counts "I've already run this exact wisp too many times — stop."

The only existing guard against repeated identical tool calls is the per-agent-loop RepetitiveToolCallDetector. It has two gaps:

1. It's rebuilt fresh per AgentLoopRunner.RunAsync and dies with that loop — so it never spans completion-eval reprompts, scheduled re-fires, or message/A2A re-triggers, which is exactly how a wisp gets re-dispatched across many short loops.
2. spawn_wisps results embed fresh batch-{Guid}/wisp-{Guid} IDs, so two byte-identical dispatches never look identical to a detector keyed on tool+args+result.

SpawnWispsExecutor is the single choke point every wisp flows through, and a process-wide singleton there is the only guard that survives across loops, reprompts, and message-driven re-triggers.

What

WispDispatchCircuitBreaker keeps a fixed-window dispatch count keyed by the exact definition hash. Once a definition exceeds DispatchCircuitBreakerMaxPerWindow (default 30) within DispatchCircuitBreakerWindow (default 5m), further dispatches of that definition are refused until the window rolls over.

Design choices:

• Exact definition hash, not the value-stripped shape hash — wisps that legitimately vary by date/id never trip it; only a truly identical re-dispatch loop does.
• Fixed window keeps memory at O(distinct definitions) regardless of dispatch rate — which matters precisely in the runaway case the breaker exists to contain. (A sliding window with per-attempt timestamps would balloon under a 400k/min loop.)
• A tripped dispatch does no work (no LLM, no tool calls) and is not written to the execution log, so a runaway collapses into a cheap, self-limiting error instead of an unbounded spin — and doesn't re-bloat the JSONL log.
• The refusal message is stable (no counts/ids), so identical refusals also look identical to the agent-loop's repetitive-result detector, which then nudges the model off the loop.
• Emits a rockbot.wisp.circuit_breaker.trips counter for alerting on runaways.

Conservative by default and fully configurable via WispOptions; the disabled flag and a non-positive limit both short-circuit to always-allow.

Tests

7 new tests (fake TimeProvider): below/at threshold, sustained trip stays tripped, window reset re-allows, per-definition independence, disabled / zero-limit / empty-hash always-allow. Full RockBot.Wisp.Tests suite green (142/142). Agent composition root builds clean.

Not included (still open from the incident write-up)

The RepetitiveToolCallDetector blind spot to GUID-bearing results, and the O(n²) FileWispExecutionLog.ReadAllAsync full re-parse on every successful wisp, are left for separate PRs.

🤖 Generated with Claude Code