diff --git a/docs/search-quality-audit-20260917.md b/docs/search-quality-audit-20260917.md
index 01974cc90..e21cf1e3a 100644
--- a/docs/search-quality-audit-20260917.md
+++ b/docs/search-quality-audit-20260917.md
@@ -75,6 +75,10 @@ Further provider inspection found that the news-to-general fallback dropped the
 
 ### Additional informal/multi-part live checks
 
+Targeted replay `reports/clean-v3-search-quality-2026-09-17T21-13-13-145Z.json`: correction-only request made zero tool calls (7.3s), but only corrected some words rather than returning the whole corrected sentence. Firefox (26.5s) now follows failed `web_fetch` with `private_browser`, proving recovery was exercised. The browser still returned a challenge title and empty snapshot; the final omitted links and was incomplete. Browser navigation success must not be conflated with successful evidence acquisition.
+
+The preceding short-news trace exposed contradictory harness controls: “no more tools” was followed twice by a demand to search again because the breadth check counted successful searches, not attempted follow-ups. Breadth recovery is now one-shot, only before a second attempt and before terminal search completion. A stream regression covers a successful first search and empty second search, preserving the final answer rather than demanding endless breadth. Related suites: 1,226 passed. Live replay remains required.
+
 `reports/clean-v3-search-quality-2026-09-17T21-09-49-686Z.json` completed five additional cases. No overall quality pass: short misspelled news took 48.9 seconds and exhausted research without synthesis; Firefox instructions took 32.2 seconds and omitted requested links; a context-free “can u look it up” invented a game-release topic; correction-only text incorrectly triggered news research. The Python false-premise answer rejected Python 9.0, but its extra latest-version claim still needs source verification.
 
 The Firefox trace showed HTTP-200 access-challenge pages treated as article evidence. `d1db1353` classifies short interstitials using corroborating title/body signals, emits an explicit fetch failure with recovery guidance, leaves ordinary articles intact, and avoids caching transient challenges. `44b56a46` preserves the supplied-text boundary for correction-only phrasing. Combined regression run: 1,249 passed. Both deployed; live targeted replay pending. Neither unit tests nor deployment establishes improved research quality.