fix: prevent active SV pacing interval from ratcheting slower - #51

Summary

Addresses the final pacing review finding from PR #50.

Change

• the nominal SV pacing interval is learned only from the initial four consistent unpaced observations;
• once active, the nominal interval remains immutable for that stream/transport session;
• moderate injection or scheduler lateness cannot feed back into the estimator and progressively reduce the wire rate;
• a deliberate sample-rate change starts a new transport/session instead of attempting to infer a new rate from already paced intervals.

Tests

Adds deterministic coverage proving that:

• long stalls do not alter the active nominal;
• repeated moderate lateness does not ratchet the nominal upward;
• even nearby paced observations are not fed back into the active estimator.

Frame encoding, sample counters, quality, APPID, VLAN, timestamps and per-stream/short injection locking remain unchanged.