feat(accounting): normalize provider evidence and savings

[alteixeira20]

Problem

Provider-reported usage, local estimates, and serialized-byte reduction are different evidence classes, and cache-token buckets do not have identical semantics across providers.

Changes

Add a small provider-neutral accounting API that:

• distinguishes provider-reported, estimated, bytes-only, and unavailable evidence;
• treats Anthropic cache read/write buckets as disjoint from uncached input;
• treats OpenAI-compatible, Featherless, and Google cached tokens as subsets of input;
• prefers provider-reported baselines over estimates;
• never turns invalid or negative counters into savings;
• keeps PXPipe-added and model latency as separately supplied measurements;
• exports the normalization primitives from the core package.

Scope

No dashboard or tracker rewiring is included.

This PR defines and tests the accounting contract only; it does not introduce a new savings claim.

Validation

Validated from a branch based directly on current upstream 0.13.1:

• production dependency audit;
• TypeScript typecheck;
• focused accounting tests;
• full test suite;
• build.

No raw prompts, credentials, session files, or machine identifiers are included.