fix(langchain): stop emitting space-separated provider names that price at $0

[viniciusdsmello]

Problem

Cost is resolved by lowercasing a step's provider and matching an llm-costs slug exactly. Matching normalizes case but not separators, so any provider value containing a space is unmatchable and the row silently prices at $0 — with no error anywhere.

Verified against the live data-stream API, identical model (deepseek-ai/DeepSeek-R1) and token counts (1000/500):

provider sent	resulting cost
Together AI (what we emit today)	0.0
Together_AI	0.0065
together_ai	0.0065

Two causes

1. The unmapped-ls_provider fallback manufactured the bug. It ran .replace("_", " ").title(), converting values LangChain already supplies as valid slugs into unpriceable display strings — vercel_ai_gateway → Vercel Ai Gateway. This silently zeroed cost for every provider not explicitly enumerated in LS_PROVIDER_TO_OPENLAYER_MAP (30 of the 139 published slugs contain an underscore; hyphen-only slugs were unaffected because .replace() only touched _). Removing the .replace() fixes them all; .title() is kept because matching is case-insensitive.

2. Six map values used literal spaces, in both LS_PROVIDER_TO_OPENLAYER_MAP and LITELLM_PREFIX_TO_PROVIDER_MAP:

• Together AI → Together_AI
• Fireworks AI → Fireworks_AI
• Hugging Face → HuggingFace

LANGCHAIN_TO_OPENLAYER_PROVIDER_MAP was already clean (all single-word) and is unchanged.

Hugging Face is not actually fixed. llm-costs publishes no huggingface provider slug at all, so that provider stays unpriced regardless of the value used. Changed for consistency only.

Verification

Live round-trip driving the patched _extract_model_info and pricing whatever provider it emits:

case	provider emitted	live cost
together_ai/deepseek-ai/DeepSeek-R1	Together_AI	0.0065 (was $0)
fireworks_ai/accounts/fireworks/models/code-llama-13b	Fireworks_AI	0.0003 (was $0)
gemini/gemini-2.5-flash (canary)	Google	0.000186 (unchanged)
anthropic/claude-3-5-sonnet-20241022 (canary)	Anthropic	0.0105 (unchanged)

tests/lib/integrations/test_langchain_callback.py: 59 passed (44 before + 15 new). Full suite shows 34 failures both before and after in deterministic serial order — no regressions; those are pre-existing (missing optional deps, plus order-dependent trace-context pollution in tests/test_tracing_core.py).

New TestProviderCostSlugs asserts no map value contains a space and that known prefixes resolve to real slugs, keeping already-correct providers as regression canaries.

⚠️ Behavior change

step.provider for an unmapped ls_provider now retains underscores: Some_New_Provider rather than Some New Provider. This also changes the derived step name ("Some_New_Provider Chat Completion"). test_ls_provider_title_cases_unknown_values was updated accordingly — it previously asserted the buggy behavior. Anything displaying provider verbatim will show the separator.

Same class of bug as OPEN-11695; worth linking a Linear ticket before merge.

[viniciusdsmello]

Added a second commit that strengthens the guard, after comparing against openlayer-ts.

openlayer-ts already solves this class of bug: langchainCallback.ts declares PROVIDER_COST_SLUG (canonical display name → cost slug, with null for vendors that have no pricing upstream) and its test suite asserts every mapped provider resolves a price. Notably it does not split the field at runtime — step.provider still gets the canonical name; the map exists purely as a test-enforced invariant. That's why every TS canonical name is single-word: multi-word ones can't price.

This PR now ports that invariant to Python. The original space check catches the specific bug but not the general class — a value can be space-free and still price nothing. test_every_provider_value_resolves_a_cost_slug asserts every value in all three maps lowercases to a slug published by llm-costs, or is an explicitly listed display-only vendor (huggingface is the only one).

Red/green verified — against pristine main it fails with:

LS_PROVIDER_TO_OPENLAYER_MAP['fireworks'] = 'Fireworks AI' lowercases to
'fireworks ai', which is neither a published cost slug nor a known-unpriced
vendor: every row would cost $0

One deliberate divergence from TS: TS sidesteps together_ai/fireworks_ai by simply not mapping them, so a TS user behind a LiteLLM proxy to those vendors gets no provider resolution at all. Python maps them and now prices them (Together_AI → together_ai), which is why underscores appear here and not in TS.

Slug set is vendored rather than fetched so the suite stays offline; refresh with:

curl -s https://llm-costs.openlayer.com/v1/costs | jq -r '.costs[].provider' | sort -u