plan_runner: fix deepening wedge + reliable email approval gate

[Bernardo1998]

Applies STEWARD_FIX.md Issues A and B (plus the D/E prerequisites that were missing in this tree) to the shared plan-runner engine. Because the scaffolder only generates per-task data + a thin run.py that imports the package, fixing these engine files means every existing and future SETUP_GUIDE-scaffolded task inherits the fixes automatically — once this lands on master (what SETUP_GUIDE §1 clones).

Issue A — deepening steps no longer wedge the plan (runner.py)

A step that passed its hard criteria but failed the soft rubric parked in deepening with no work queued, so select_next_action returned stuck every cycle, forever.

• needs_deepening now generates deepening work for plan_step dispatches and advances the parent budget for deepening dispatches.
• New promotion pass before _advance_stages flips spent/empty-queue deepening steps to done_enough, so the stage gate arms instead of wedging.

Issue D — /approve marker was never detected (context.py, plan_phases.py)

check_stage_approval/check_intent_answer read only feedback["body"], but g10_parse_feedback stores text under raw_reply and body was never set.

• load_context mirrors the reply into feedback["body"]; check_stage_approval also reads raw_reply for robustness.

Issue E — replies were never fetched (runner.py, feedback.py)

_check_email_replies early-returns unless last_email_date is set, and the runner never set it.

• Persist last_email_date + subject_prefix into state after a sent email and thread them via meta_for_context.
• send_ltt_email falls back to global_state["subject_prefix"] before [LTT], so a plan_runner task's sent subject and its reply IMAP search use the same prefix (they previously mismatched → replies never found).

Issue B — reliable email gate (plan_phases.py, context.py, runner.py)

• New interpret_reply_nl: conservative LLM reader for plain-English approvals; no-ops unless a gate is armed.
• Plumb the reply Message-ID into feedback["reply_msgid"].
• One-reply-one-gate de-dup (last_reply_approval_msgid) + nl_reply fallback at cycle start, plus same-cycle approval after _advance_stages to close the arm-at-end / read-at-start timing gap.

scaffolder.py: add last_email_date, subject_prefix, last_reply_approval_msgid to the default state template.

Not touched: Issue C (already correct here), Issue F (out of scope per decision).

Verification

• 180/180 existing tests pass; py_compile clean on all changed files.
• Issue A: reproduced the stuck wedge, confirmed fix → done_enough → gate armed.
• Issue D: /approve detected via body and raw_reply; wrong-stage tokens don't jump ahead.
• Issue B de-dup: one reply approves at most one gate; a fresh Message-ID approves the next.
• Issue B live LLM smoke test: "approve, proceed with your tests." → approve; "hold on, let me check the numbers first" → hold.
• Scaffold + verify of a fresh task succeeds with the new state keys.

🤖 Generated with Claude Code