fix(providers): normalize OpenAI-compatible cache usage

[Raymond8196]

Summary

Parse prompt-cache usage through one shared normalizer across every OpenAI-compatible path (non-streaming, streaming, SSE reassembly), so persisted cache usage, context-size, and hit-rate metrics stay consistent regardless of which vendor cache dialect the provider returns.

Why this is needed. The OpenAI-compatible client is shared by OpenAI, Zhipu (GLM), DashScope, Groq, xAI, MiniMax, DeepSeek, and aggregator relays (OpenRouter, etc.). Each reports cache hits in a different shape, but the three execution paths previously only forwarded raw prompt_tokens / completion_tokens / total_tokens — so cache counters were silently dropped and hit_rate read as a constant 0% on providers that do report cache.

Shared normalizer (Usage::to_usage_map) keys on field schema (not provider id) and resolves three real-world dialects by priority:

Upstream shape	Example fields	Normalized
Standard nested (OpenAI / Zhipu / DashScope / Groq / xAI / MiniMax)	prompt_tokens incl. prompt_tokens_details.cached_tokens	prompt = raw − cache_read, emits cache_read
OpenRouter cache write	nested cache_write_tokens	prompt = raw − cache_read − cache_write, emits cache_write
DeepSeek hit/miss split	prompt_cache_hit_tokens + prompt_cache_miss_tokens	miss → prompt, hit → cache_read (missing half derived from inclusive total)
Already-normalized relay	top-level cache_read_tokens / cache_write_tokens (prompt excludes cache)	passed through, no subtraction

Tie-break rule: a populated nested shape wins; a non-empty DeepSeek split wins over an empty {} nested; already-normalized top-level counters are only used when neither vendor shape is present. All fields accept camelCase aliases. Saturating subtraction guards against malformed counts producing negative prompt.

Contract kept cache-exclusive. prompt_tokens is normalized to uncached/billable input; cache_read and cache_write live in their own counters. Downstream UsageTotals reconstructs the real context-window fill (last_prompt = prompt + cache_read + cache_write) so a fully-cached prefix no longer reports as "2 tokens used", and cache_hit_rate = cache_read / (cache_read + prompt) reflects actual cost savings.

Test plan

• cargo test -p agent_core --lib providers::openai_compat: 78 passed / 0 failed, covering:
  • standard nested cached_tokens, OpenRouter nested cache_write_tokens
  • DeepSeek hit+miss split, hit-only / miss-only derivation from inclusive total, empty-nested-vs-split tie-break
  • camelCase aliases, plain usage (no cache fields), top-level already-normalized counters
  • malformed nested count saturates prompt to 0 (never negative)
  • end-to-end wiremock: OpenAI non-streaming, DeepSeek streaming, SSE reassembly
• cargo clippy -p agent_core --lib: 0 warnings in the 5 changed files
• Production-verified on live GLM-5.2 traffic: per-call hit_rate 95–98% on warm rounds; cache_read genuinely non-zero (60224 / 58432 / 59968 …), confirming the nested cached_tokens shape is captured end-to-end.