fix: node distance interpolation for flipped/single reaches

[jameshgrn]

Summary

• Fix inverted node distances on 660 flow-corrected reaches (offset reversal for stale v17b node dist_out)
• Fix single-node reach placement from upstream edge to centroid (25,732 nodes)
• Sync v17c_flow_corrections table: add 31 missing flipped IDs, remove 7 oscillation-reverted false positives (636 → 660)
• Add 3 node distance columns (hydro_dist_out, hydro_dist_hw, dist_out_dijkstra) to NetCDF export spec and schema migration
• Fix flow_verification.py::rebuild_derived_attrs missing flipped_reach_ids and update_node_columns call
• Fix 3 pre-existing test failures (stale is_mainstem_edge, missing topology_suspect column)
• Update release notes: 5 → 8 node-level v17c variables
• Organize export files into versioned directories with version-tagged filenames and fresh SHA256 checksums

Validation

• 0 negative distances across 5 columns, 11.1M nodes
• 0 NULLs across 8 propagated node columns
• 0 intra-reach monotonicity violations (660 flipped + 248K normal)
• 716 tests pass, 163 skipped, 0 failed (full suite)
• ruff check + format clean
• All 54 export files (0.0.2 + 0.0.3 + 0.0.4) version-tagged and checksum-verified

Export file organization

Reorganized data/exports/ into versioned directories with version-tagged filenames:

data/exports/
├── 0.0.2/netcdf/   6 files  *_v17c_0.0.2.nc         (6/6 SHA256 OK)
├── 0.0.3/          24 files  *_v17c_0.0.3.*          (24/24 SHA256 OK)
│   ├── netcdf/     6 NC
│   ├── gpkg/       6 GPKG
│   └── parquet/   12 Parquet
└── 0.0.4/          24 files  *_v17c_0.0.4.*          (24/24 SHA256 OK)
    ├── netcdf/     6 NC
    ├── gpkg/       6 GPKG
    └── parquet/   12 Parquet

Updated /preflight skill check #9 (export version tagging) to enforce versioned directory + filename convention.

Test plan

• 5 new interpolation regression tests (normal/flipped/single-node/isolation/no-negatives)
• Full test suite: 716 passed
• DB preflight checks 1–5 pass (node ordering, monotonicity, propagation, ghost exclusion, hydro_dist_hw freshness)
• Previously-failing tests now pass (column_order, flow_direction, flow_verification)
• 0.0.4 exports: 0 negative distances, 0 NULLs across 6 regions
• All 54 export files version-tagged with fresh SHA256 checksums verified

[jameshgrn]

Code Review Summary

Findings: 0 P1, 0 P2, 1 P3, 3 P4

P3 — Fixed

• Dry-run display off-by-one (fix_node_distances.py:105): row[:6] included node_id, shifting all variable assignments. Fixed to unpack all 7 columns with _nid discard. Display-only — no data impact.

P4 — Informational (no action)

1. db_path parameter unused in rebuild_derived_attrs (pre-existing)
2. SQL IN clause via string interpolation — pragmatic for DuckDB with 800+ IDs from trusted source
3. Good dedup: single interpolation source in output.py, maintenance script imports it

Verification

• 23/23 output tests pass, 46/46 flow tests pass
• ruff check + format clean
• Interpolation math spot-checked (flipped offset, hydro_dist_hw, hydro_dist_out)
• flow_verification.py compute order matches v17c_pipeline.py

Fix pushed as f7f7550.