#820 Report when the connector's backlog warns, pauses and recovers

Summary

Make a growing connector backlog visible even when no reporting callbacks are installed. Warn, pause, resume, and recovery transitions are logged with their depth and thresholds.

Warnings and pauses clear only after the backlog drains to half their thresholds, avoiding a stream of repeated messages around a boundary. Pausing describes a producer waiting for room, not necessarily the live feed: retries and recovery walks also offer work. The held depth is the capacity observed by the blocked offer, not a counter that could include other in-flight offers.

Verification

• Full local bin/ci passes on d7d55874bcde21bc4f3cb68ca9591e7eb5da7d6f.
• Five focused queue/reporting regressions pass over 20 race-enabled runs.
• Tests check the actual log levels, transition order, depths, thresholds, recovery bands, reporting before callbacks, fluctuating workloads, and concurrent offers. Callback-delivery assertions wait for the callback rather than guessing its scheduling.
• Existing queue behavior and cancellation tests remain in the normal suite. No new harness, fake Basecamp, or testing framework.

The panic-containment contract remains callback-specific; this PR does not introduce general fault tolerance around application logging handlers.