Fix review feedback: CRS guard, logging, upgrade loop

[NewGraphEnvironment]

Summary

Fixes from Claude Code review of #19:

• Guard src.crs.to_epsg() against None CRS (returns epsg=None instead of AttributeError)
• Log failed metadata extractions with warning instead of silent swallow
• Fix infinite re-extraction loop: detect old-format rows by transform column (not epsg) so files with valid metadata but no EPSG are not re-extracted every run
• Accurate docstring: two remote reads per URL during extraction (rasterio + cog_validate)

Test plan

• 3-item extraction + creation works
• Second run fully cached, zero re-extractions

🤖 Generated with Claude Code

[claude]

Review

Changes are correct and fix the three issues cleanly.

Upgrade loop fix (item_create.py:156-161): The transform column is the right sentinel. On a successful extraction, transform is always a non-null JSON string even when epsg=None. On a failed extraction, transform=None (stored as NaN in CSV). So pd.isna(row.get('transform')) correctly distinguishes 'never extracted' from 'extracted but no CRS'. The previous epsg-based check conflated those two states.

One edge case to be aware of: Line 158 checks row.get('is_geotiff'). Pandas reads CSV boolean columns as bool by default in recent versions, but if is_geotiff loads as the string 'True' due to mixed-type inference, the truthiness check still works correctly — not a bug, just worth knowing.

CRS guard (stac_utils.py:113): Correct. src.crs can be None for GeoTIFFs missing a CRS definition; the guard prevents AttributeError.

Logging (stac_utils.py:131-132): Better than silent swallow. No issues.

No bugs introduced. LGTM.