The once-per-window warning map kept every distinct error forever; an
error whose text varies per tick would grow the daemon without bound.
Entries older than the window are now dropped before each check.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dsh4Znwh2KnC7q6vbGpX1N
observe_shadow caught only OSError/ValueError, so a RecursionError from
leisure_watch escaped into the unguarded root loop and crash-looped the
daemon, stopping enforced billing. It now catches Exception, fails closed
(failclosed:watcher, unless the pause check itself answered "paused") and
logs each distinct error at most once per 10 minutes.
resolve_budget priced the same answers twice (enforced + leisure), so
earned_time's "No answer for the X earner" warning fired twice per tick.
The shadow pricing now gets the answers completed with None, which
earned_time treats identically to missing, so only the first warns.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dsh4Znwh2KnC7q6vbGpX1N