When a secure-listener writer hit its bounded send-credit stall deadline, the mux already reset the stream and reclaimed capacity, so recovery worked. The diagnostic path did not: it stayed INFO-level and carried no measured age, so diagnosis did not work and an owner could not identify the wedge from logs alone. The new stall_age_ms field measures the current contiguous zero-credit wait, not stream lifetime and not cumulative prior stalls. The origin is cleared only when credit is actually consumed, not on every wake. Because a zero-credit WINDOW wakes the waiter and restarts the wait_for deadline without ending the zero-credit condition, a reported age can legitimately exceed the 120s deadline, and that excess is the wedge signal an owner needs. Only the starvation context is owner-warning severity. Every other reset context stays INFO and carries no age, so the new WARNING is a specific, greppable signal rather than added log volume. The deadline, RESET_CANCEL reason, send_credit_starvation context string, framing, admission, and client behavior are unchanged. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>