V023: 2 new pathlen_out step violations from routing weight change

[jameshgrn]

Summary

The learned routing weights (commit a676073) introduced 2 new pathlen_out_step violations (V023). These did not exist in the pre-routing database (sword_v17c_pre_routing.duckdb).

Affected reaches

reach_id	Region	pathlen_out	dn_pathlen_out	Step	Expected	River
81210700081	NA	491,138	535,421	-44,282	11,043	Deer Hunting Slough
52139900545	OC	51,093	73,152	-22,059	6,602	NODATA

Root cause

At these bifurcations, the new weighted score produces different rankings when evaluated in different contexts:

• compute_best_headwater_outlet (downstream pass) picks successor A → computes pathlen_out along that path
• compute_main_neighbors picks successor B → rch_id_dn_main points to B
• Result: pathlen_out increases when following rch_id_dn_main

This is a routing divergence between the two functions at close-scoring bifurcations.

Impact

0.0008% of checked edges (2 / 244,739). Non-blocking for beta.

Options

1. Accept as INFO — downgrade V023 threshold or add these 2 reaches to a known-exception list
2. Investigate the bifurcations — check if _up_key double-count fix (this branch) resolves them
3. Tighten consistency — ensure both functions see identical score at these junctions (may require passing the same pathlen context)

[jameshgrn]

Fixed. Root cause: compute_main_neighbors used mainstem chain at bifurcations, but the chain was constrained to within-group successors. At bifurcations where the best successor is in a different main_path_id group (e.g. Deer Hunting Slough vs Yukon River), the chain forced the wrong pick. Fix: at bifurcations (n_rch_down > 1), use score-based ranking instead of chain. Chain remains authoritative on 1:1 links. V023 now 0 violations globally.