feat(relay): read conversation history from agent transcripts

An agent TUI runs on the alternate screen, so herdr keeps no scrollback
for it (max_offset_from_bottom is 0 on every agent pane here), and the one
read that does reach older rows costs ~31ms per line and visibly scrolls
the operator's terminal. History now comes from the transcript the agent
writes for itself, keyed by the agent_session ref on the pane record. The
session uuid never crosses the wire: clients send a pane_id and the relay
resolves it through pane_session_map.

relay/transcript.py locates the file by session uuid (a glob, 0.7ms --
deriving the path from the pane's cwd is wrong, because claude fixes its
project directory at its own startup cwd and the two drift), parses
claude's JSONL into turns, pages backwards from the newest, and caches
parsed turns rather than raw bytes (33MB of JSONL is 0.25MB of turns;
cold 227ms, warm 1ms). Remote panes are read over SSH in one round trip
using only ls/wc/tail/head, and the relay hands the far side the size it
already has, so paging a remote pane usually moves no bytes -- at the
cost of remote history being recency-bounded, which file_truncated says
out loud.

Two things the real files forced:

- rows get replayed (591 of 2602 in one session here, same uuid, same
  content), so they are deduped by row uuid; otherwise every turn renders
  twice and a turn uuid cannot serve as a pagination cursor
- 200-turn pages measured 97KB to 324KB of JSON, so a page is bounded by
  characters as well as by turns, with the remainder reachable through
  has_more

The panes the relay reports now carry title, scrollback and has_session,
which is what lets a client know a pane has a transcript to ask for.