0.5.4

Fixed

• A model participant that is rate limited now waits as long as the endpoint asks (Retry-After, the x-ratelimit-reset-* buckets, or a retry hint in the body) instead of retrying within seconds against a per-minute limit and burning the rest of the quota (#58). Being throttled is no longer a failure and no longer marks a participant unavailable, so a limit that clears in a minute does not pause it for ten (#44); each turn that ends throttled doubles the wait instead, up to fifteen minutes, and any answer resets it. The cooldown belongs to the profile, so every room using it holds back until it clears, and GET /api/health reports rate_limited and rate_limited_until separately from failures.

Added

• rate_limit_backoff_ms on a model profile (default 20000): the first step of the doubling wait used when an endpoint refuses without saying how long to wait.

Images (multi-arch):

• ghcr.io/dev-dull/maindmeld:0.5.4@sha256:36fa4fb40c291f3edb81b81648ed9243c5c969d346e48b042fc88056d738cf43
• ghcr.io/dev-dull/maindmeld-runner:0.5.4@sha256:f92edc0d8f775a7082510c230564ca1a09dd1b3752262d782798d2b19c022881

Note: the v0.5.4 tag's tree carries the version bump but not the CHANGELOG heading or the image tags in README and compose; those landed in 7850f6e on main right after. The published images are built from this tag and are unaffected.