RUNBOOK: record the rejected multilingual-e5 trial with its numbers

RUNBOOK.md

@@ -115,10 +115,16 @@ git push mirror --all
   detection) was considered and deliberately skipped: ~6 minutes of
   rate-limited CanLII API per run for marginal signal. Run `canlex_case` on
   a decision before relying on it instead.
-- French support is phase 1 only (accent-folded tokens + FR→EN synonym
-  bridges; the `french` eval slice tracks it). Phase 2 = multilingual
-  embedder via the `CANLEX_EMB_REPO` env hooks; phase 3 = parallel French
-  corpus.
+- French support is phase 1 (accent-folded tokens + FR→EN synonym bridges;
+  `french` eval slice: Hit@3 0.80). Phase 2 — swapping to
+  multilingual-e5-small via the `CANLEX_EMB_REPO` env hooks — was TRIED AND
+  REJECTED 2026-07-22: it regressed every slice (legacy MRR 0.87→0.82,
+  french Hit@3 0.80→0.60, since the synonym bridges already anglicize the
+  query) and broke confidence calibration (no-answer probes 0.67→0.84,
+  above the 0.72 hedge). Don't retry without also re-calibrating
+  HEDGE_THRESHOLD and adding untranslatable French queries to the eval.
+  Phase 3 (parallel French corpus from the bilingual Justice Laws XML)
+  remains the real path to full bilingual support.
 - Typo robustness: unknown query tokens get a trigram-Dice fuzzy bridge
   (0.6 floor, shared 2-char prefix, 0.7 score discount) in
   `index._fuzzy_term`; the `typo` eval slice tracks it.