fix(split-pane): open Pane B's live queue after the response, keep the disconnected marker last

[timkjr]

Follow-ups from the #506 review.

• The history pull's live-frame queue opened before the fetch, so Pane B stopped painting for the whole round trip. It now opens when the response arrives, beside capturedAt. Frames from before it are replaced by the capture's \x1bc or written unchanged, so the final buffer is the same. The request uses the primary pane's CodemanFetchDeadline budget, and the body read gets 10 s once headers land, since that is the only window that holds live output.
• A {t:'r'} refresh queued behind a pull ran its clear() after the marker was written and wiped it. A close during a refresh load also wrote the marker above the replay. The marker is now an owed flag (_markerOwed) that each load settles in its own finally, so it is the last thing on screen however the close, the pull and a trailing refresh interleave.

Two related items in one PR because they share the marker/queue state machine; happy to split if preferred.

Testing: unit tests drive each ordering. Checked in real Chrome against a beta instance with a delayed pull, old code vs new:

Scenario	Old code	New code
Output while the fetch is in flight	held	painted
Close mid-pull	marker last	marker last
Close mid-pull with a refresh queued	marker wiped	marker last

docs/architecture-invariants.md updated.

🤖 Generated with Claude Code