bug: Streaming responses record 0 actual tokens — cost and quota tracking broken for all streams

[stackbilt-admin]

Bug

Every streaming response is recorded with zero input and output tokens. Cost and quota tracking are entirely inaccurate for streamed requests.

Location: src/factory.ts:1064

// buildFactoryStream() — called for all streaming responses
const usage = { inputTokens: 0, outputTokens: 0, totalTokens: 0, cost: estimatedCost };

The token fields are hardcoded to 0. Only the pre-request cost estimate (derived from character count) is stored. No post-stream reconciliation happens.

Downstream effects:

• CreditLedger spend is based on rough estimates, not actual usage — monthly budgets can drift significantly
• CostTracker per-provider totals are wrong for any provider used via streaming
• QuotaHook receives actualCost: estimatedCost with no token breakdown (src/factory.ts:1079)

Root cause

generateResponseStream() returns a ReadableStream<string> rather than an LLMResponse, so the token counts in the final SSE event are never extracted.

Providers do surface final usage in their SSE streams:

• Anthropic: message_delta event with usage.output_tokens; message_start with usage.input_tokens
• OpenAI / Groq / Cerebras: final data: chunk with usage.prompt_tokens / usage.completion_tokens when stream_options: { include_usage: true } is set

Fix

1. Pass stream_options: { include_usage: true } in OpenAI/Groq/Cerebras streaming requests
2. Buffer the final SSE chunk in each provider's stream path and extract usage
3. After the stream drains, call the same recordQuota / costTracker.trackCost path used by non-streaming responses

Acceptance criteria

• Streaming responses record actual inputTokens and outputTokens post-stream
• CreditLedger and CostTracker reflect real token counts for streamed requests
• Existing streaming tests updated; new tests cover token accumulation across providers

Found by

Codebase audit (automated) — src/factory.ts:1047–1090