Bug: google.ts input double-counts cached tokens (inconsistent with google-gemini-cli.ts)

[reverendrewind]

Bug

In packages/ai/src/providers/google.ts, usage metadata maps promptTokenCount directly to input:

input: chunk.usageMetadata.promptTokenCount || 0,
cacheRead: chunk.usageMetadata.cachedContentTokenCount || 0,

Per Google's documentation, promptTokenCount includes cachedContentTokenCount as a subset. So when a downstream consumer computes input + cacheRead, cached tokens are counted twice.

Real data from a cached request:

• promptTokenCount = 21675, cachedContentTokenCount = 16289
• Mapped: input = 21675, cacheRead = 16289
• input + cacheRead = 37964 — nearly double the actual totalTokenCount of 21680

This causes false context overflow detection in OpenClaw (openclaw/openclaw#15265).

Fix

The correct implementation already exists in google-gemini-cli.ts (line 660):

const promptTokens = responseData.usageMetadata.promptTokenCount || 0;
const cacheReadTokens = responseData.usageMetadata.cachedContentTokenCount || 0;
input: promptTokens - cacheReadTokens,

Apply the same subtraction to google.ts. One-line change, I have a branch ready to go once approved.