Every gate was green — and stopped one request short

The 400 only exists on the second request of a run that calls a tool and loops. No pre-ship check ever made a second request.

Before deploy
pass
model registered
config check — no live request
request depth
—
pass
unit tests
offline — no live request
request depth
—
pass
single-turn smoke call
proves the provider accepts one request — acceptance, not runtime
request depth
turn 1
deploy
In prod
never exercised
turn-2 tool-call replay
the only place the bug can fire
fails in prod at 06:30
request depth
turn 2 — where the bug lives
scroll →
Every existing gate stopped one request short of where the bug lives: the turn-2 replay was never exercised until production.