feat(cost): config-level rates override for models missing from _RATES

[Joi Ito (Joi)]

Problem

Models absent from the hardcoded _RATES table in _cost.py make compute_cost() return None, so every usage.cost_usd for those models is null. Today this affects all claude-fable-* sessions: in my logs, 129 of 129 fable responses have cost_usd: null. microsoft/amplifier#303 documents the family-detection side of the same gap. Any private, fine-tuned, or newly released model hits this until a code change ships.

Solution

A new rates provider config key supplies rates for models the built-in table does not know:

providers:
  - module: provider-anthropic
    config:
      rates:
        claude-fable-5:
          input: 10.00        # USD per 1M tokens, same units as _RATES
          output: 50.00
          cache_read: 1.00    # optional, defaults to 10% of input
          cache_write: 12.50  # optional, defaults to 125% of input
        claude-fable-*:       # trailing-* glob covers dated ids
          input: 10.00
          output: 50.00

• Lookup order: exact override, longest matching trailing-* glob override, then _RATES. Config rates take precedence over the built-in table.
• Cache-field defaults (10% / 125% of input) are the ratios every existing _RATES row uses.
• Values are parsed via Decimal(str(x)). No float arithmetic touches the cost path.
• None still means unknown model with no rate anywhere. Unknown is never silently zero.
• Invalid entries are skipped with a warning, matching the provider's lenient config parsing (_config_int / _config_float style). They never fail the mount.
• The speed='fast' 2x multiplier logic is unchanged and still applies only to _FAST_ELIGIBLE_MODELS.

_RATES itself is untouched.

Relationship to #58

Complements it. #58 adds fable/mythos rows to _RATES and should still land for out-of-box fable pricing. This PR covers the general case: any model missing from the table, without waiting for a release. The override row shape and the fable example rates here match #58.

Tests

tests/test_rates_override.py adds 21 tests: parse validation (missing/invalid/unknown fields, non-mapping input), derived cache defaults, Decimal precision (no float artifacts), override hit for an unlisted model, precedence over _RATES, unknown model still None, glob matching (exact beats glob, longest prefix wins), fast-multiplier interaction, and three integration tests through _convert_to_chat_response with config-supplied rates.

Local run (uv, amplifier-core 1.6.0 from PyPI): 499 passed, 1 skipped, and 10 pre-existing failures/errors that reproduce identically on clean main in the same environment (test_tool_repair, test_validation, test_behavioral, test_cost_usd_json_serializable::test_usage_model_stores_decimal_internally). No regressions from this change; all cost tests (test_cost.py + new file, 41 tests) pass.

🤖 Generated with Claude Code