fix(eval): use uuid session ids, extract prompt text, and drop structural telemetry on deep learn

[theonly1me]

What changed

• Use valid UUIDs for Claude Code replay session IDs in runEval so sessions are not rejected by the CLI.
• Extract user prompt text from raw JSON transcript envelopes in extractPromptText.
• Capture stderr in runClaudeCode and surface engine errors if an eval run fails.
• Drop raw structural telemetry templates from the compiled profile when shadowclone learn --deep runs.
• Move score dimension averaging into src/eval/score.ts to keep src/eval/run.ts within the line limit.

Why

Claude Code exited immediately on non-UUID replay session IDs and transcript JSON strings were passed directly as prompts, causing eval to report zero delta without running. In addition, deep distillation was concatenating coarse structural templates with semantic rules instead of replacing them.

How to verify

bun run check
bun test src/eval/prompt.test.ts
bun test src/eval/run.test.ts