chore: record the news-scorer reactive-coverage cap in the live test deviations

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01B3PXXn2tHBzdv6rYtGPz86

research/live_test_freeze.json

@@ -90,6 +90,11 @@
       "what": "TrueNAS ollama2 app turned off by the owner (host was out of memory: 843 MiB free of 15 GiB; tools:latest held 7.8 GB at a 131k context). Settings unchanged, but the LLM catalyst score (NEWS_CATALYST_SCORE_ENABLED) returns nothing while it is off, so the rank boost is inert. The dilution veto falls back to its regex floor.",
       "restore": "Owner chose qwen3:14b on mason instead of re-enabling TrueNAS Ollama (2026-10-01). config_overrides set OLLAMA_HOST=http://192.168.1.178:11434, OLLAMA_MODEL=qwen3:14b, OLLAMA_THINK=false (needs image >= eed4ec0). Takes effect at the next live-stack restart; set 'to' to that time. The scoring model is then qwen3:14b, not the frozen tools:latest. That is a known, recorded change to the news-scoring input.",
       "outcome": "News scoring restored with qwen3:14b on mason (OLLAMA_THINK=false, image 4361f80). The first restart (22:22) ran the OLD image (TrueNAS had not pulled), and 320 empty-answer rows were cached as 'other'; they were deleted after the container was recreated at 22:36. From 22:37 the scorer is qwen3:14b, not the frozen tools:latest; known weakness: reactive articles and halt-up notices can score as bullish catalysts."
+    },
+    {
+      "from": "2026-10-02 10:10 ET",
+      "to": null,
+      "what": "News scorer prompt + deterministic cap (3e523e1): halt notices and after-the-fact 'why is X up' / movers coverage capped at conviction 0.1 / 0.3, so they no longer give a catalyst rank boost. Real catalysts and dilution flags unchanged (20-headline check: reactive mean boost 0.318 -> 0.055, real 0.353 -> 0.362). Deployed by a restart DURING market hours (10:10 ET) with TNMG open; its Tradier bracket stop (4.43) stayed in place and the held-symbol guard blocks a re-entry."
     }
   ]
 }