fix(ai): report OpenAI cache writes - #38735

What

Report GPT-5.6 and later OpenAI prompt-cache writes in normalized AI usage for both Responses and Chat Completions.

GPT-5.6 cache writes are billed at 1.25x uncached input, but the native protocol decoders currently discard cache_write_tokens. This makes every write appear as zero and overstates fresh input.

Adds a patch changeset for @opencode-ai/ai.

Before / After

Before

Given provider usage with 5 total input tokens, 1 cache-read token, and 2 cache-write tokens, normalized usage reported:

input:       5
fresh:       4
cache read:  1
cache write: undefined

After

The same payload preserves the provider's complete non-overlapping breakdown:

input:       5
fresh:       2
cache read:  1
cache write: 2

The invariant fresh + cache read + cache write = input remains true.

How

• packages/ai/src/protocols/open-responses.ts decodes input_tokens_details.cache_write_tokens and maps it into Usage.cacheWriteInputTokens.
• packages/ai/src/protocols/openai-chat.ts does the same for prompt_tokens_details.cache_write_tokens.
• Both protocols subtract reads and writes when deriving nonCachedInputTokens from the provider's inclusive input total.
• Existing stream-fixture tests now cover the complete GPT-5.6 usage shape.

Scope

This PR changes response accounting only. It does not alter request bodies, enable GPT-5.6 explicit cache breakpoints, or change automatic cache policy.

Testing

• bun run test test/provider/openai-responses.test.ts test/provider/openai-chat.test.ts: 116 passed, 0 failed
• bun run test in packages/ai: 444 passed, 28 skipped, 0 failed
• bun typecheck in packages/ai: passed
• pre-push bun turbo typecheck --concurrency=3: 33 tasks passed
• git diff --check: passed