Keep mid-turn stalls and call counters off the transcript

[justrach]

The line REPL was reading as a receipt of the harness, not the task.

A typical turn printed ⚠ stream stalled while tools were still running, then · turn still going · model call 4 · 3 tools, then [stream stalled — reconnecting (SSE) (2/2)]. Those lines sit at the same weight as › Wrote 1 file. People hit Esc on a turn that was still recovering.

After this, a turn looks like the work:

#10 › Are you able to open it up for me?
  › Searched 1 pattern
  › Wrote 1 file
  I'll write docs/showcase.html and open it in Chrome.

What changed:

• Mid-turn stall/drop (the reconnect ladder) is silent on both the line REPL and the pager. Tracer and telemetry still record stream_retry.
• The inner-loop model call N · N tools pulse is gone. Tool rows already show the work.
• A ⚠ still prints when the turn actually ends (stream stalled — ending turn / connection dropped — response ended early).
• ■ interrupted, ↳ queued, and ↳ force stay — those are your actions.

This is ADR 0021 compliance (transcript is the task, not the event bus), not a new decision. Stalls still reconnect (WS, then SSE). Opening Chrome in five model calls is a separate capability gap.

Test plan

• Mid-turn transport_aborted emits no notice (line sink + TUI sink)
• Ending-turn stall/drop still prints the ⚠
• Give-up after FORCE_STALL_ALWAYS / FORCE_DROP_ALWAYS prints the ending-turn notice (scripts/test-stall-drop.py)
• agent_request no longer says the reconnect line; stream_retry stays on the tracer
• formatTurnPulse is empty; beforeRequest does not emitNotice
• scripts/eval-tier1.sh unit/reach/fmt/lines green (suite 1756)

  

[justrach]

Landed on release/v0.0.281 (#670) as merge 7fa82ee. Closing this vs main so it does not double-land; it ships when 281 merges.