fix(magnitude-core): fix cumulative token counting in ModelHarness._reportUsage()

The _reportUsage() method was incorrectly overwriting prevTotalInputTokens
and prevTotalOutputTokens instead of accumulating them. This caused the
tokensUsed event to emit inflated token counts (10-15x higher than actual
API usage) when using non-Anthropic providers (e.g., openai-generic).

Changed assignment operators from = to += to properly track cumulative
token usage across multiple LLM calls.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Author: 2 people

packages/magnitude-core/src/ai/modelHarness.ts

@@ -170,8 +170,8 @@ export class ModelHarness {
         this.events.emit('tokensUsed', usage);
         //console.log("Usage:", usage);
 
-        this.prevTotalInputTokens = inputTokens;
-        this.prevTotalOutputTokens = outputTokens;
+        this.prevTotalInputTokens += inputTokens;
+        this.prevTotalOutputTokens += outputTokens;
     }
 
     async partialAct<T>(