Salvages PR #19160 (@liquidchen) and adds an auto-pause guard so weak judge models can't silently burn the entire goal turn budget.

Changes

• cherry-picked a9f5f81 from PR fix(gateway): defer goal status notices until after response delivery #19160 (contributor authorship preserved via rebase-merge): routes goal status notices through adapter.send() instead of a non-existent send_message(), defers the notice until after the main response is delivered, and cancels queued synthetic goal continuations on /goal pause and /goal clear
• new: GoalState.consecutive_parse_failures counter + DEFAULT_MAX_CONSECUTIVE_PARSE_FAILURES=3 threshold
• _parse_judge_response() / judge_goal() now return a 3-tuple (verdict, reason, parse_failed) — parse_failed is True only on empty/non-JSON judge output, False on API/transport errors (those stay transient, fail-open)
• evaluate_after_turn() auto-pauses after 3 consecutive parse failures with a config-pointer message:

  ⏸ Goal paused — the judge model (3 turns) isn't returning the required JSON verdict.
  Route the judge to a stricter model in ~/.hermes/config.yaml:
    auxiliary:
      goal_judge:
        provider: openrouter
        model: google/gemini-3-flash-preview
  Then /goal resume to continue.
  

• test fix: test_goal_verdict_send.py was sharing a hardcoded session_id="goal-sess-1" across tests. It only worked before because _post_turn_goal_continuation was a never-awaited coroutine — the coroutine never ran, so cross-test state never leaked. PR fix(gateway): defer goal status notices until after response delivery #19160's async conversion made the coroutine actually run, surfacing the latent xdist test-leakage bug. Each test now gets a unique session_id via uuid.uuid4().
• AUTHOR_MAP entry: ytchen0719@gmail.com → liquidchen

Why

Discord user barteq reported an infinite /goal loop when using deepseek-v4-flash as the judge. The judge returned output like:

judge returned empty response
judge reply was not JSON: "Let me analyze whether the goal is fully satisfied..."

The loop fail-opened to "continue" on every such turn, burning the entire 20-turn budget.

Gille already has a config-side workaround (route auxiliary.goal_judge to google/gemini-3-flash-preview), which is the right answer long-term — but this PR makes the agent detect the failure mode and tell the user HOW to fix it, instead of running silently until the budget is exhausted.

PR #19160 (same session) was also needed because /goal clear didn't stop an already-queued synthetic continuation — related failure mode, same shape.

Validation

E2E tested via execute_code with real imports from the worktree, isolated HERMES_HOME, and mock judge clients returning the exact shapes barteq reported.

Credit

• @liquidchen — PR fix(gateway): defer goal status notices until after response delivery #19160 cherry-pick (gateway send()/defer/continuation cleanup + 3 regression tests)
• Teknium — Option B direction (auto-pause with config pointer)

Closes #19160.