Log GenPlan progress and preserve failure diagnostics - #173

Summary

• Log GenPlan request/validation timing, costs, stopping, and training progress to the console and sandbox/genplan.log.
• Save validation feedback and failed-request diagnostics, and write the attempted prompt before calling the model.
• Surface captured Claude CLI stdout/stderr on errors and log held-out evaluation progress.
• Add regression tests for progress logging, saved failure artifacts, and CLI diagnostics.

Verification

89 tests passed across GenPlan, best-of-K, validation, LLM clients, and prompts; 2 live integration tests excluded. git diff --check passed. No live model calls.

Scope

No retry logic, budget changes, token accounting, or source-bundle changes. uv.lock, campaigns, and results are excluded.

This is the base PR for source-access fix #172.