fix: prevent deferred inference requests from starving the event loop

[antra-tess]

Summary

• always yield a real event-loop turn when inference requests are deferred behind a busy agent
• preserve low-latency processing when queue traffic and background work coexist
• throttle stale-request traces together with their once-per-minute warning
• add regression coverage for the exact scheduler states observed during the Sol outage

Root cause

When a new wake arrived while an agent was streaming or waiting_for_tools, processInferenceRequests() requeued it. processNextEvent() treated the non-empty request queue as progress and skipped its normal backoff. If activeStreams was empty or stale at the same time, the loop had no macrotask yield at all.

That microtask-only polling loop pinned Bun at 100% CPU and starved the tool-result and HTTP I/O needed to make progress. Sol remained waiting_for_tools, while Fleet health probes timed out and reported the agent down.

Impact

Deferred wakes can no longer turn an otherwise recoverable tool wait into a self-sustaining event-loop wedge. Tool results, Web UI health checks, Discord traffic, and timers continue receiving event-loop time while the wake remains queued.

Validation

• npm run build
• focused scheduler, reconnect, and channel-routing tests
• full agent-framework suite: 326 passed, 0 failed
• deployed to Sol as a canary: Fleet returned UP, Web UI HTTP 200, zero pending requests, and normal CPU