README: name the class of bug that mattered most

Three separate bugs each billed the user, reported success and returned something
broken. None was a crash, and none showed up in a benchmark — only in end-to-end
runs. Worth saying plainly at the top level rather than only in RECIPES.md.

README.md

@@ -99,6 +99,12 @@ Three results worth knowing, all measured rather than assumed:
   1 of 20 answers short); from 1e-4 up it works and the value barely matters. The
   literature's rate clears it with ~2x margin, and lowering it "to be safe" would
   break the product silently.
+- **Silent failures are the real enemy.** Three separate bugs each produced a run
+  that billed the user, reported success, and returned something broken: a merged
+  gpt-oss missing 23 of 24 layers' experts, a repo left holding a previous run's
+  stale artifact, and a run diverged to NaN that pushed a model answering every
+  prompt with "!!!!". None was a crash. All three are guarded now, and all three were
+  found by running the thing end to end rather than benchmarking it.
 
 ## The endpoint