Reclassify 78 midstream ghost reaches (type=6 with both up and down neighbors) #199

Problem

78 reaches globally are labeled type=6 (ghost) but have topology neighbors in both directions (n_rch_up > 0 AND n_rch_down > 0). Ghost reaches should only exist at network boundaries (headwater or outlet), not midstream.

These are misclassified — they carry real data (width, WSE, facc, slope) and sit between real reaches. Many are tiny slivers (62–790 m long, median ~200 m).

Distribution

Characteristics

• 67/78 have end_reach=0 (midstream)
• 27/78 have lakeflag=1 (lake) — likely lake/river boundary classification artifacts
• All have valid width, facc, dist_out
• Named rivers include Amazon (3,060 m wide!), Wisconsin, Grijalva, Elk, Mattagami, Savannah, Spree

Impact

• These break path_segs contiguity (N016 lint check) — a path_segs=-9999 ghost in the middle of a segment splits it into two disconnected fragments. All 51 global N016 violations trace back to midstream ghosts.
• Every lint check needs type NOT IN (5, 6) exclusions to avoid false positives from these.

Detection

SELECT reach_id, region, river_name, width, reach_length, lakeflag, end_reach
FROM reaches
WHERE type = 6 AND n_rch_up > 0 AND n_rch_down > 0
ORDER BY region, river_name

Also available via workflow.find_incorrect_ghost_reaches().

Proposed fix

Reclassify based on lakeflag:

• lakeflag=0 → type=1 (river)
• lakeflag=1 → type=3 (lake_on_river)
• lakeflag=2 → type=1 (canal, keep as river type)

After reclassification, re-run v17c pipeline on affected regions to recompute path_freq, stream_order, path_segs for the formerly-ghost reaches.