[thorwhalen]

Closes #81.

Adds falaw.llm_call_cost_usd(provider, model, usage, *, prices=None) -> (usd, known), one first-party price table for Anthropic and OpenAI (falaw/data/first_party_llm_prices.json, every row with its own source and date), plus get_llm_price, list_llm_prices, llm_prices_table_version, LlmPrice.

• Rows: the union of reelee agent.ANTHROPIC_PRICES_USD_PER_MTOK (12 models, Haiku 5.5 long-prompt rates, cache ratios materialized per row), reelee chat_relay.OPENAI_PRICES_USD_PER_MTOK (8 models), and foley's Anthropic subset (contained in reelee's). Dates carried over: Anthropic 2026-10-06, OpenAI 2026-09-28. I did not re-verify the numbers against the vendors; they are the copies' numbers.
• usage: a dict or SDK object; Anthropic's four token fields; OpenAI's Responses and Chat Completions shapes (cached tokens at the cached rate, input includes cached).
• Unlisted model: the provider's dearest listed rates and known=False. This is not a ceiling (an older, pricier model can cost more), so a caller that must never under-count, like reelee's chat relay, which refuses an unpriced model, has to refuse when known is false.
• Exact model ids only, as in reelee's relay (a dated snapshot is not its family).
• Backward compatible: purely additive; nothing existing changes. llm_rates (fal-routed LLMs) is untouched.

Verification: new tests (tests/test_llm_prices.py), doctests, full suite 696 passed. A one-off cross-check against reelee's model_call_cost_usd and price_usd on 3,000 random usages (all models plus unknown ones, all cache fields, both sides of the 100K Haiku threshold) gave 0 mismatches.

reelee / foley need: a falaw release, then a falaw bump. reelee agent.model_call_cost_usd and chat_relay.price_usd can become thin wrappers over llm_call_cost_usd (keeping their import paths); foley _usage_cost / _estimate likewise. I did not touch either repo.

Adversarial review (fresh Opus subagent): problems found, all confirmed and fixed here. (1) token counts that were floats, numeric strings or numpy ints priced at $0 with known=True: now parsed like reelee's int(...); (2) the unknown-model "upper bound" was only the dearest listed rate, so wording corrected (and the migration note above); (3) 1-hour cache writes were charged at the 5-minute rate: new cache_write_1h_usd_per_mtok column (2x input), read from usage.cache_creation.ephemeral_1h_input_tokens; (4) with a caller-supplied table the unknown-model bound ignored long-prompt rates: now included. Also: cached tokens beyond input are no longer capped (matches reelee, the higher number), non-string model ids no longer crash, the loaded table is read-only. Re-ran the 3,000-case cross-check against reelee after the fixes: 0 mismatches.