docs: add NDVI batch debugging as pre-requisite in spec

Generated with [Claude Code](https://claude.ai/code)
via [Happy](https://happy.engineering)

Co-Authored-By: Claude <noreply@anthropic.com>
Co-Authored-By: Happy <yesreply@happy.engineering>

docs/superpowers/specs/2026-04-01-batch-sar-buildup-water-design.md

@@ -152,9 +152,25 @@ For each indicator, add a test mirroring the existing NDVI batch test pattern:
 
 Use `rasterio.MemoryFile` to create synthetic GeoTIFFs for test fixtures, matching each indicator's expected band layout (interleaved VV/VH for SAR, single-index bands for buildup/water).
 
+## Pre-requisite: Debug NDVI Batch
+
+NDVI batch currently returns placeholder results despite having `uses_batch = True`. This must be fixed before adding batch to the other three indicators — otherwise we'd be replicating a broken pattern.
+
+**Debugging steps**:
+1. Add targeted logging to the worker's batch phases: log the exact exception when `submit_batch()` fails, log job statuses during polling, log download errors during harvest
+2. Test CDSE connection locally — verify `get_connection()` succeeds with the configured `OPENEO_CLIENT_ID` / `OPENEO_CLIENT_SECRET`
+3. If auth works, submit a minimal batch job manually and track it through submit → poll → download
+4. Fix whatever is broken (auth config, connection handling, download path issues, etc.)
+5. Verify NDVI produces `data_source="satellite"` results end-to-end before proceeding
+
+**Possible root causes** (ordered by likelihood):
+- CDSE credentials not set or invalid on HF Spaces (memory note: "CDSE creds added, untested")
+- `get_connection()` falls through to browser-based OIDC which fails non-interactively
+- Batch jobs submit but error on CDSE side (graph too large, quota exceeded)
+- Download paths not found after `download_results()` (the `_find_tif` issue)
+
 ## Out of Scope
 
-- Debugging why NDVI batch currently returns placeholder results (likely CDSE auth — tracked separately)
 - Changes to graph builders in `openeo_client.py`
 - Changes to worker orchestration or polling logic
 - New shared abstractions or base class methods