fix(dashboard): parse CodeBuddy index.json for token usage

[jeff-r2026]

Problem

Token/prompt usage statistics never populated for CodeBuddy sessions (always 0), while Claude sessions worked.

Root cause: token stats have a single source — scanTranscriptStop() scanning the transcript_path at Stop time. That scanner is hard-bound to the Claude Code JSONL schema (line-by-line JSON.parse of {type:"assistant", message.usage.*_tokens}). CodeBuddy's Stop hook does fire and does pass a transcript_path, but it points to a single pretty-printed index.json document (CodeBuddy's own format), so every line fails JSON.parse and tokens stay 0.

CodeBuddy's index.json actually contains authoritative usage:

{
  "messages": [{ "role": "user" | "assistant" | "tool" }],
  "requests": [{ "usage": { "inputTokens", "outputTokens", "totalTokens" } }]
}

Fix

Add scanCodebuddyIndex():

• tokens: sum requests[].usage.{inputTokens,outputTokens} (same per-turn accumulation model as the Claude scan; input + output matches CodeBuddy's own totalTokens). CodeBuddy exposes no cache-read/creation split at the request level, so those map to 0.
• prompts: count messages[] with role === "user".

Dispatched from scanTranscriptStop() when the transcript basename is index.json, with a safe fallback to the existing Claude JSONL scanner (returns null on any parse failure). No change to the Claude path.

Verification

Tested end-to-end against a real local CodeBuddy transcript:

• scanTranscriptStop(index.json) → {input: 215609, output: 2967}, prompts: 2 (total 218576 matches the file's totalTokens)
• Full parseHookEvent(Stop payload, "codebuddy") path attaches tokens + prompts
• Compiled artifact via the real hook CLI command dashboard-report --stdin --tool codebuddy writes the tokens into events.jsonl

Added a regression test; vitest (dashboard-collector + conversation-token-metrics, 83 tests) and tsc --noEmit pass.