Usage tracking, model-tier feedback loops, and agent-coordination signals

Failure modes

Symptom	Cause	Fix	Severity
`get_stats` totals are 0	No calls recorded yet, or buffered and unflushed	Run a workflow, or call `flush()`	low
`recommend_tier` keeps returning the current tier	Fewer than `MIN_SAMPLES` (10) feedback rows for the stage's tier	Record more feedback before expecting a switch	medium
`recommend_tier` says "No feedback data available" despite recorded feedback	Tier recorded with non-lowercase casing (e.g. `"CAPABLE"`); it only matches `cheap`/`capable`/`premium`	Record with lowercase tier strings	medium
Coordination calls fail / no agents listed	Redis-backed features unavailable	Check `TelemetryFeatures.is_redis_available()`; install the Redis extra	medium
Stats look stale	Entries older than retention/TTL aged out	`retention_days` (90) bounds the store; feedback `FEEDBACK_TTL` is 7 days	low

Buffered writes. UsageTracker batches; a reader may miss the newest calls until flush() or the buffer fills (buffer_size).
MIN_SAMPLES gate. Tier recommendations are conservative by design — they need 10 samples before they move.
Redis-gated coordination. EventStreamer / HeartbeatCoordinator need Redis; guard with TelemetryFeatures.

UsageTracker.get_instance().get_stats(days=1) — is anything recorded?
flush() then re-read, to rule out buffering.
For tier recommendations, get_quality_stats(...) returns None only when the stage has no feedback; recommend_tier needs MIN_SAMPLES (and lowercase tier strings) before it moves.
For coordination, TelemetryFeatures().is_redis_available().