The guard was described as covering the bedrock cache-token regression
(#25846). It does not. That bug needs prompt_tokens to exclude cache tokens
so prompt_tokens - cache_hit falls below zero; this test drives uncached
traffic, so cache_hit is 0 and the subtraction cannot go negative. Reverting
the clamp leaves the test green, confirmed by mutation: the only way to turn
it red was an injected fault unrelated to the bug.
State the scope and the limitation in the docstring, mark the cell
fail_before_fix=unproven explicitly, and record the live bedrock cache-token
path as a still-open gap in the matrix. The test keeps its real value as a
cross-provider non-negative invariant sweep.
e2fb677f3c