[szemyd]

This PR is stacked on #31 and merges into its branch. It fixes review findings checked against the server in dream-faster/unravel-router#1149 and dream-faster/unravel-router#1199.

What changes

• A drop during a catch-up resumes where the replay got to. The live rows held during a replay arrive before replayed, so lastBar runs ahead of a replay that is still catching up. A reconnect used to resume from it and skip the rest of the gap. replayFrom now tracks each unfinished replay's bar, advanced by each replay frame's time, and a reconnect resumes from there first.
• An empty truncated replay (to null) is asked for again from from. If it comes back empty again from the same bar, that bar is larger than the server's cap. Stderr then says so instead of "more to come", and the stream stops asking. StreamReplay.FollowedUp tells the CLI whether a follow-up was sent.
• A refused follow-up is sent again. A follow-up refused with limit_exceeded is sent again after a jittered pause, up to streamBusyRetries times, then the stream warns. Running out of the 50 follow-ups also warns now, instead of stopping without a word.
• Docs: the streamMessageLimit comment is updated (replayed bars now come in small frames), and so is the README's replay paragraph.

Tests

go test -run 'TestStream_' ./... passes. New tests:

• an empty truncated replay is asked for again from from, then given up with a message;
• a drop during a catch-up resumes from where the replay got to;
• a refused follow-up is sent again, with no warning.

go vet and gofmt are clean. The full go test ./... has the same 8 failures as this PR's base branch: they are the *_live_test.go suites, which need APERIODIC_API_KEY and network access.

https://claude.ai/code/session_01MoDcuxddwYnzAPetT7cR5r

---

Generated by Claude Code