fix(todos): teach the priority marker in the machine's own authoring hint

The planner offers the current priority band, so an Agent Todo whose text has no
`[P0]/[P1]/[P2]` marker is written to the active state and then cannot be bound
with `quota should-run --todo-id` in the same Turn. Reproduced on the
2026-09-17 and 2026-09-20 wakes of goal `loopx-meta`: the rows authored from the
hint's own `agent_todo_command_template` answered
`state=quota_action_selection_rejected reason=candidate_not_currently_eligible`,
and seven already-open rows in that lane had to be repaired by hand before the
planner could offer them.

The hint is the machine-readable guidance every heartbeat reads, so its agent
template now prints the marker: `--text '[P1] <agent action>'`. The rule sentence
is deliberately unchanged — agent-facing CLI output has a measured per-row
budget (the first draft added a second field and failed
`cli-output-budget-regression-smoke` with +271 chars against a 128-char
allowance), and a template the agent copies teaches the marker without growing
every quota should-run row.

Validation: new tests/control_plane/test_todo_write_hint_contract.py (template
carries the marker; rule and user templates unchanged),
`project-agent-adoption-smoke` ok with the added assertion,
`cli-output-budget-regression-smoke` ok, 65 quota settlement CLI tests passed,
ruff and mypy clean, and `loopx canary premerge --from-git-diff` passed with 0
failures and 0 manual holds.

Signed-off-by: huangruiteng <14976749+huangruiteng@users.noreply.github.com>

Author: huangruiteng

examples/project/project-agent-adoption-smoke.py

@@ -256,6 +256,10 @@ def main() -> int:
         assert approved_quota["todo_write_hint"]["agent_todo_command_template"].startswith(
             f"loopx todo add --goal-id {GOAL_ID} --role agent "
         ), approved_quota
+        # The planner offers the current priority band, so the machine's own
+        # authoring template has to teach the marker that makes a row bindable.
+        assert "[P1]" in approved_quota["todo_write_hint"]["agent_todo_command_template"], approved_quota
+        assert "[P1]" in approved_quota["todo_write_hint"]["agent_todo_command_template"], approved_quota
 
         packet = run_cli(root, registry_path, "review-packet", "--goal-id", GOAL_ID, "--scan-root", str(project))
         assert packet["ok"] is True, packet

loopx/control_plane/todos/write_hint.py

@@ -5,6 +5,13 @@
 
 def build_todo_write_hint(goal_id: str) -> dict[str, str]:
     return {
+        # Agent Todos are offered to the planner by the current priority band, so
+        # a row authored without a marker is written but cannot be bound by the
+        # same Turn. Following this hint on 2026-09-17 and 2026-09-20 produced
+        # open, claimed rows that `quota should-run --todo-id` then rejected with
+        # `candidate_not_currently_eligible`. Only the template teaches the
+        # marker: agent-facing CLI output is budgeted per row, and this hint has
+        # no room to grow.
         "rule": "Write user/owner actions to User Todo, not Next Action/docs/chat.",
         "user_gate_command_template": (
             f"loopx todo add --goal-id {goal_id} --role user "
@@ -16,7 +23,7 @@ def build_todo_write_hint(goal_id: str) -> dict[str, str]:
             "--task-class user_action --bound-agent <id> --text '<action>'"
         ),
         "agent_todo_command_template": (
-            f"loopx todo add --goal-id {goal_id} --role agent --text '<agent action>'"
+            f"loopx todo add --goal-id {goal_id} --role agent --text '[P1] <agent action>'"
         ),
         "section": "User Todo / Owner Review Reading Queue",
     }

tests/control_plane/test_todo_write_hint_contract.py

@@ -0,0 +1,52 @@
+"""The machine's own Todo-authoring hint must teach the bindable row shape."""
+
+from __future__ import annotations
+
+from loopx.control_plane.todos.write_hint import build_todo_write_hint
+
+
+def test_agent_template_carries_the_priority_marker_the_planner_requires() -> None:
+    """A row authored from this template must be bindable in the same Turn.
+
+    The planner offers the current priority band, so an Agent Todo whose text
+    has no ``[P0]/[P1]/[P2]`` marker is written to the active state and then
+    cannot be bound with ``quota should-run --todo-id``. Reproduced on the
+    2026-09-17 and 2026-09-20 wakes; the hint has to name the marker instead of
+    letting an agent copy a template that produces an unselectable row.
+    """
+
+    hint = build_todo_write_hint("fixture-goal")
+    template = hint["agent_todo_command_template"]
+    assert template.startswith(
+        "loopx todo add --goal-id fixture-goal --role agent "
+    ), hint
+    assert "--text '[P1] " in template, hint
+    # The rule sentence is unchanged: agent-facing CLI output is measured per
+    # row, so the marker is taught by the template the agent copies.
+    assert hint["rule"] == (
+        "Write user/owner actions to User Todo, not Next Action/docs/chat."
+    ), hint
+
+
+def test_user_templates_are_not_repointed_at_the_agent_band_rule() -> None:
+    """User rows keep their own contract; only the Agent template changes.
+
+    The marker guidance rides in the existing ``rule`` sentence instead of a new
+    field: agent-facing CLI output has its own measured budget, and a separate
+    field doubled the growth of every quota should-run row.
+    """
+
+    hint = build_todo_write_hint("fixture-goal")
+    assert hint["user_gate_command_template"] == (
+        "loopx todo add --goal-id fixture-goal --role user "
+        "--task-class user_gate --blocks-agent <agent-id> "
+        "--text '<blocking user decision>'"
+    ), hint
+    assert hint["user_action_command_template"] == (
+        "loopx todo add --goal-id fixture-goal --role user "
+        "--task-class user_action --bound-agent <id> --text '<action>'"
+    ), hint
+    assert hint["section"] == "User Todo / Owner Review Reading Queue", hint
+    assert hint["rule"] == (
+        "Write user/owner actions to User Todo, not Next Action/docs/chat."
+    ), hint