#1920 r16: CLAUDE.md contradicts the usage-less classification

[realmarcin]

Codex round 16 on PR #1920 (head 5c53659), should-fix. The sentence says an observation without terminal usage stays recovered, then says it is transcript_observation_invalid, never recovered; the latter belongs to malformed and overlapping observations only. Fix in #1920.

[realmarcin]

Fixed in 4152c1c on PR #1920: CLAUDE.md says a usage-less observation stays recovered with the counter shown, and separately that a refused observation is reported as transcript_observation_invalid. Closes when #1920 merges.