Phase 1: extract the climb loop's decide step into a composition seam

[renmengye]

What

Phase 1 of the research-loop buildout (docs/design/research-loop-buildout.md, merged in #127): extract the climb loop's decide-next step into a pluggable composition seam.

climb_once's inner loop is already run → measure → decide. run (run_role) and measure (measure_and_decide) were factored; the "decide the next invocation" half was inlined in the loop. This pulls it out as a pure policy:

• _panel_revise_policy(verdict, session, harness, reads, revisions) -> _Revise | _Halt
• the loop now just runs the policy: _Halt ends (drafting if findings stay open), _Revise wakes the author with the policy's prompt and re-enters run → measure.

Why

Depth (Phase 2a) is a results-driven decision policy; parallel (Phase 3) fans out the run step. With the decision behind a seam, they become apps on the loop, not new drivers — the whole point of Phase 1, and the antidote to the per-lane-orchestration mistake (#123).

No behavior change

The one policy today is the existing panel-driven revision rule, making byte-for-byte the same decisions (clean → halt; blocking + resumable + under cap → revise; unresumable or capped → halt-and-draft). This is deliberately thin — run_role consolidation was already done (consolidation.md), so Phase 1 is only this decide-seam extraction.

Test plan

• New test_panel_revise_policy_decisions covers each decision branch directly (the seam).
• The full existing climb/orchestrator suite stays green (no behavior change): uv run pytest — 697 passed. Full gate (ruff check + format, mypy) — green.

No secrets / no large files

Confirmed.

🤖 Generated with Claude Code

[github-actions]

Round 1 — reviewed head 91040698 — reviewer hermes/gpt-5.6-terra.

second opinion — terra
Advisory findings from autoresearch — the code owner decides. Reply to disagree; the autoresearch:no-review label opts this PR out.

Verdict: no defects found.

[github-actions]

Round 1 — reviewed head 91040698 — reviewer claude/claude-opus-5.

Advisory findings from autoresearch — the code owner decides. Reply to disagree; the autoresearch:no-review label opts this PR out.

Verdict: nothing blocking — 1 advisory note.

Advisory (non-blocking):

• Docstring names a credited argument the function does not take (src/autoresearch/orchestrator.py:957; medium)

I compared the extracted policy against the deleted inline branches: the order and conditions are identical (not blocking -> halt; no session_id or supports_resume False -> halt+draft; reads > revisions -> halt+draft; else revise with verdict.wake_text), and the call passes panel_reads after the increment, matching the old code. The one behavioral difference is that the clean-read halt now assigns panel_blocking_open = False instead of leaving it untouched, but panel_blocking_open is initialized False at line 1085 and every place that sets it True breaks immediately, so the value cannot differ. The new test exercises all four branches with distinct inputs and would fail if a branch were inverted. No other files changed in the diff; I did not run the suite (no execute tool).