Fixed real bugs found by live testing this session:
- The fallback state machine could misread "nothing has committed yet" as "the user navigated away," aborting recovery when an instance hung or failed at the network level (offline, DNS failure, connection refused) instead of falling back to the next one.
- A stale record could get clobbered by a newer navigation mid-fallback.
- The redirect-loop breaker no longer counts Nitter-to-Nitter fallback hops, so a multi-instance failure cascade doesn't suppress interception for 30s.
- Firefox's speculative/predictive connections (requests with no real tab) no longer crash the fallback logic.
Instance list: xcancel.com received the same cease-and-desist as nitter.net and shut down. lightbrd.com is disabled independently - it doesn't proxy media and loads third-party analytics (zedeus/nitter#1209). Down to 2 active instances (catsarch.com, kareem.one); the public Nitter fleet is under broader pressure from X Corp right now, see README.
CI: fixed a bug where a commented-out instance was still counted as shipped, made a real shipped/manifest mismatch fail the job, pinned web-ext for reproducible lint.