The test lowered AITokenLimit to 8000 to force truncation, but a
terminal-linked chat registers ~25 tool schemas costing ~5,500 tokens,
which maxTokens subtracts from the budget alongside the conversation.
The system prompt plus one padded round then pushed even turn 1 over
8000, so every turn was rejected pre-send with requestTooLarge and no
blob was ever stored (blobCount 0 != 6).

Raise AITokenLimit to 16000 so the tool-schema floor plus a round fits,
and raise AIResponseTokenLimit to 7000 so it clears that floor too:
truncate() budgets maxTotalTokens - maxResponseTokens and counts only
the conversation, not the tool schemas, so unless maxResponseTokens
exceeds the tool floor the request is rejected before truncation can
activate. With both raised, turn 1 fits, later rounds cross the
truncate threshold, and the oldest round's blob is dropped from the
final request as the test asserts.