Estimate cost+duration before running full image analysis

Run a single cheap vision pass to estimate the cost+duration of the full from-image flow. Free to the user; we eat ~$0.02/call internally.

The frontend uses this output to render a "Will cost $X-Y, balance $Z. Proceed?" confirmation before committing to the expensive perception path. Auto-mode can skip the confirmation; customize-mode shows per-asset choices once we have asset-level data (P2).