Tracking telemetry

What the tracking pipeline reports about itself: provider health, the click channel's own lifecycle, and the attention decisions that were committed and refused. It is written after decisions, never before, by a recorder that owns nothing — the deciding is AttentionEngine's (AttentionOrderSpecs.md). Switching telemetry off cannot change what AltTab does.

The QA harness is the consumer: --qa-state reads summary, and --qa-telemetry drains the ring as NDJSON.

Record schema

• every record carries v, seq, at and kind; every other field is optional and omitted when unset
• sequence numbers are monotonic and never reused, so a drained batch stitches onto the previous one
• the field set is closed and carries no window title, keystroke or document name
• NDJSON lines sort their keys, so two runs diff cleanly

Ring buffer

• records accumulate until drained
• the ring is bounded; the oldest records are dropped first and droppedCount reports the gap
• draining returns the buffered records and empties the ring

Attention

• a committed decision records the pid, wid, process generation, source, reason code and status
• source is the CHANNEL that named the window (workspace, altTab, accessibility, annotatedSession), so a disagreement is attributable to one of them rather than to "focus" in general
• lastAttention always reflects the most recent COMMITTED decision — a refusal is recorded as its own event rather than overwriting it, because a reader asking "where does the model think the user is" is not helped by the answer it just declined

AX provider health

• health is tracked per pid, with lifecycle state, observer generation, attempts, capabilities and last error
• an observer generation change is a rebuild: attempts and last error start again
• a process exit drops its entry

Session tap

• installed/enabled are lifecycle facts; decoded and invalid counts are per event
• an event that failed to decode is counted, never guessed at