#1979 S8: probe classifier: Meta's max_output_tokens floor wording; 402 Payment Required → unavailable (billing) #1996

Parent: #1979. Found in the hub live test of PR #1995, 2026-10-28.

• meta (6 models): \max_output_tokens` The number must be `>= 16`.The S7 floor matcher only knows "Expected a value >= N", so these come back aserrorand are hidden. Broaden it to the general shape (max_output_tokens / max_tokens / maxOutputTokens near a>= N, minimum (value)? (is)? N`).
• baseten (14 models): HTTP 402 Payment Required: please check your current payment status. is classified as error. It's an account or billing state, so classify it unavailable with the reason "billing: payment required". Treat "exceeded your current quota … billing" the same way: that's billing, not a transient rate limit. A plain 429 stays rate-limited.

AC (RED first):

• The Meta floor message triggers the retry at 16.
• 402 → unavailable with a billing reason.
• Google's quota/billing message → unavailable.
• A plain 429 'rate limit' → rate-limited (unchanged).
• Existing tests still pass.