fix: emit user-visible DEGRADED RUN WARNING on bare named-entity calls

The stderr [Planner] warning from PR #285 doesn't reach the user because
Claude and other reasoning agents hide stderr from their synthesis. The
2026-04-19 Hermes Agent Use Cases Run 1 produced source=deterministic
and the user never saw it.

Adds a user-visible stdout block that the model's LAW 5 pass-through
contract forces into the response. Fires only when plan_source is
deterministic AND no pre-research flags were passed AND the topic is
pre-research-eligible (named entity). Cron jobs on abstract topics
don't trigger it.

Position: BEFORE the EVIDENCE FOR SYNTHESIS envelope so the model sees
it as the first non-badge content. Wrapped in a new USER-VISIBLE BANNER
envelope matching the EVIDENCE/PASS-THROUGH envelope pattern from Unit 1
of PR #285.

Runtime-agnostic language: explicitly enumerates Claude Code, Codex,
Hermes, Gemini so the hosting reasoning model recognizes itself
regardless of runtime.

pipeline.py now persists plan_source to report.artifacts so the
renderer can consume it. Adds 7 tests covering fire conditions,
suppression conditions (external/llm plan source, flags present,
abstract topic), and correct position relative to the evidence envelope.

scripts/lib/pipeline.py

@@ -254,6 +254,10 @@ def run(
         print("[Planner]   (no subqueries in plan)", file=sys.stderr)
 
     bundle = schema.RetrievalBundle(artifacts={"grounding": []})
+    # Expose plan_source to the renderer so render_compact can emit the
+    # DEGRADED RUN banner when a named-entity topic was invoked bare
+    # (source=deterministic AND no pre-research flags). LAW 7 backstop.
+    bundle.artifacts["plan_source"] = plan_source
 
     # Project-mode or person-mode GitHub: run once before the main subquery loop
     _github_custom_done = False