[Agent Loop] Streaming token/cost budgets (maxToolTokens/maxCostUsd on generateStream) #854

Part of #612.

Enforce maxToolTokens / maxCostUsd on the streaming auto path (generateStream). Today these throw 501 at stream entry; the sync path enforces them.

Why deferred

GenerateChunk doesn't expose usage in v1, so cumulative spend isn't observable from a stream. Wiring this needs the chunk type extended plus each in-tree backend (openai / anthropic / bedrock) emitting usage on its final chunk.

Acceptance

• GenerateChunk (or a terminal metadata chunk) carries usage.
• Backends populate it on the stream's final chunk.
• Streaming loop tallies cumulative usage and trips BudgetExceededError like the sync path.
• Tests: streaming token-budget trip and cost-budget trip.

Note: the maxToolIterations cap already applies to streaming; this closes the spend-budget gap only.

🤖 Generated with Claude Code