Refactor usage models to Pydantic with graceful degradation

[alexkroman]

Replace imperative helper functions with declarative Pydantic models for parsing AMS usage data, improving type safety and maintainability while preserving graceful degradation on malformed input.

Summary

This refactor introduces three Pydantic models (_LineItem, _Window, _Usage) to replace seven standalone functions that parsed and formatted AMS usage responses. The models are deliberately tolerant of schema drift: unknown fields are ignored, missing/invalid prices default to 0.0, and non-list/non-object values degrade to empty collections rather than crashing. This allows assembly usage to render partial data instead of failing entirely on unexpected API responses.

Key changes

• New Pydantic models (_LineItem, _Window, _Usage):

  • _LineItem: wraps price (cents) with a label property that searches for product/feature keys in preferred order
  • _Window: aggregates line items, computes total_cents from prices, formats label from timestamps, and generates breakdown (per-product spend summary)
  • _Usage: top-level container for usage_items (list of windows)
  • All use BeforeValidator to coerce raw AMS dicts through jsonshape helpers, and extra="allow" to tolerate unknown fields

• Removed functions: _usage_items, _window_total_cents, _window_label, _line_item_name, _line_items_summary — their logic is now properties/methods on the models

• Updated call sites in body() render function to use model validation and property access instead of function calls

• Refactored config_builder.py coercion to use Pydantic's TypeAdapter validators instead of custom parsing functions, reducing duplication and leveraging Pydantic's standard error handling

• Updated tests to use model constructors (_Window.model_validate(), _LineItem.model_validate()) and assert on properties rather than function returns; added explicit test for graceful degradation on junk shapes

Notable details

• The _MappingList validator is reused across models to handle the AMS pattern of non-list/non-object values degrading to []
• Price defaults to 0.0 via BeforeValidator(jsonshape.as_float), so missing/invalid prices don't break aggregation
• The breakdown property remains a string (not a structured type) to match the existing table rendering; it's computed on-demand from line items
• --json output is unaffected: the raw AMS dict is passed through untouched before model validation

https://claude.ai/code/session_01CbHTsyv1Z4nPgH4EEHbPpM