fix(OPEN-11695): recognize LangChain Gemini provider/model and populate cost + usageDetails

[viniciusdsmello]

Problem

LangChain traces reported $0 cost with empty costDetails/usageDetails. The customer (Globo, Python SDK + ChatGoogleGenerativeAI) saw provider: chat-google-generative-ai, model: models/gemini-3.5-flash, costDetails: {}. The cost table keys on (provider, model), so an unrecognized provider or model silently yields $0.

Closes OPEN-11695.

Context after rebase

Rebased onto main, which now includes the v1 callback modernization (OPEN-11315). That work resolves the provider from metadata["ls_provider"] (so ChatGoogleGenerativeAI → Google when ls_provider is present) and captures granular token_details into step metadata. This PR closes the gaps that still produced $0 / empty cost fields on top of it.

Fix

1. Provider anchor for the exact reported string. Map the chat-google-generative-ai _type → Google in the legacy _type map, so the string named in this issue still resolves when metadata["ls_provider"] is absent (older langchain-google-genai, or callers that don't forward it) — not only via the ls_provider path.

2. Strip the Gemini models/ prefix in _extract_model_info — ChatGoogleGenerativeAI reports models/gemini-3.5-flash, which the cost table (bare names) doesn't match, so the lookup misses → $0. Runs before the existing LiteLLM prefix handling; models is never a provider, so it's safe.

3. Emit a priced usageDetails map so the backend populates costDetails. Main only surfaced token_details as informational metadata; this adds the priced column. LangChain reports cache/audio categories as overlapping subsets of the input/output totals, so they're re-partitioned to be non-overlapping and mapped to the cost table's keys:

   • cache_read → cached_tokens, cache_creation → cache_creation_tokens
   • input audio → audio_input_tokens, output audio → audio_output_tokens
   • reasoning stays folded into output_tokens (billed at the output rate; no separate price).

   ChatCompletionStep gains an optional usage_details field, serialized as usageDetails only when set — other integrations are unaffected.

Verification

• 44 tests in tests/lib/integrations/test_langchain_callback.py (main's + those added here), all passing; ruff clean. Guarded by pytest.importorskip("langchain_core") (langchain is an optional dep). Includes a regression test that chat-google-generative-ai resolves to Google with ls_provider absent.

• End-to-end against a live pipeline with the customer's inputs (27131 prompt / 17739 completion):

  field	before	after
  provider	chat-google-generative-ai	Google
  model	models/gemini-3.5-flash	gemini-3.5-flash
  cost	$0	0.2003475
  costDetails	{}	{input_tokens, output_tokens} populated

• With caching (10 000 cached input tokens): usageDetails = {input_tokens: 17131, cached_tokens: 10000, output_tokens: 17739} (non-overlapping, sums to total), costDetails prices cached_tokens at the cheaper rate, cost = 0.1868475 — correctly below the non-cached 0.2003475. A cross-provider audit confirms the cost table uses these category keys uniformly (anthropic/openai/bedrock/azure/… all use cached_tokens/cache_creation_tokens/audio_*).

Scope / follow-up

• SDK-only: helps future traces from updated SDKs; does not backfill existing $0 rows.
• The cost table is LiteLLM-fragmented across google/gemini/vertex_ai; Google + bare model resolves the reported case and most current Gemini models but not every variant — a general backend normalize-before-lookup layer would close that fully. The same $0 symptom also appears on the Langfuse/OTLP ingestion path, untouched here. Both are candidate backend companion tickets.

🤖 Generated with Claude Code