feat(http): per-turn cost_usd in /v1/chat/completions usage

[manojp99]

Summary

Surfaces per-turn dollar cost on the chat-completions HTTP face by lifting the cost_usd value that provider modules already emit on the NDJSON wire (via amplifier_agent_lib/bundle/hook_streaming.py) into the OpenAI usage envelope.

Pure wire translation — no new pricing data, no provider-side changes, no parallel catalog surface. The cost telemetry already flows on the JSON-RPC wire; this PR just plumbs it through to the OpenAI-shape wire that opencode and other OpenAI-compatible clients see.

What the wire looks like now

"usage": {
  "prompt_tokens":           9276,
  "completion_tokens":       54,
  "total_tokens":            9330,
  "prompt_tokens_details":   { "cached_tokens": 0 },
  "cost_usd":                "0.0118625"
}

cost_usd is a string to preserve Decimal precision. Standard OpenAI clients ignore the non-standard field; cost-aware clients can render the real dollar value.

Implementation

File	Change
_event_translator.py	extract_usage widens return type to dict[str, Any]; reads event['cost'] (set by hook_streaming from kernel cost_usd) and stamps cost_usd: str(...) on the result
_wire.py	_build_usage_block, stop_chunk, tool_calls_stop_chunk accept `cost_usd: str
routes/chat_completions.py	Accumulates `usage_cost: Decimal

cost_usd is omitted from the response entirely when no provider emitted it (older provider modules, third-party endpoints without cost telemetry).

Verified end-to-end

$ curl -X POST /v1/chat/completions -d '{"model":"claude-haiku-4-5-...", ...}' | grep usage
"usage":{"prompt_tokens":9276,"completion_tokens":54,"total_tokens":9330,
         "prompt_tokens_details":{"cached_tokens":0},"cost_usd":"0.0118625"}

Real $$ from Anthropic's pricing, accumulated and surfaced on the OpenAI wire.

Compatibility

• Additive — standard OpenAI clients ignore the new cost_usd field.
• No wire-protocol bump.
• No provider-side change required — the field is only emitted when the provider already supplies cost_usd (anthropic, openai, and chat-completions providers all do today).

Companion PRs

• microsoft/amplifier-module-provider-anthropic#59 — closed (catalog approach was redundant with existing NDJSON flow)
• microsoft/amplifier-app-opencode#2 — surfaces per-model limit in the opencode config (was always available in /v1/models; was stripped for paranoia)