fix: refine token usage reporting

The old token UI mixed context pressure, cache counters, and dollar estimates.

That made the percentage look precise even when provider data was incomplete or different.

The composer and /cost view now show concrete counts instead of a pie percentage.

Token payloads now share a smaller shape: used, inputTokens, outputTokens, and breakdown.

Claude uses per-step usage where available and Codex reads total_token_usage events.

Gemini reads its tokens object without inventing a context window.

OpenCode reads opencode.db session totals and includes all token columns in used.

The /cost backend no longer returns cache display fields or input/output dollar estimates.

This avoids derived values that look reliable but are not comparable across providers.

Verification: npm run typecheck; targeted eslint; OpenCode session provider test.

server/modules/providers/list/gemini/gemini-sessions.provider.ts

@@ -88,22 +88,15 @@ function buildGeminiTokenUsage(tokens: unknown): AnyRecord | undefined {
   const record = tokens as AnyRecord;
   const input = Number(record.input || 0);
   const output = Number(record.output || 0);
-  const cached = Number(record.cached || 0);
-  const thoughts = Number(record.thoughts || 0);
-  const tool = Number(record.tool || 0);
-
-  const totalFromFields = input + output + cached + thoughts + tool;
-  const total = Number(record.total || totalFromFields || 0);
+  const total = Number(record.total || input + output || 0);
 
   return {
     used: total,
-    total: total,
+    inputTokens: input,
+    outputTokens: output,
     breakdown: {
       input,
       output,
-      cached,
-      thoughts,
-      tool,
     },
   };
 }