lnk_habitat_validate() scores mad watershed groups as if they were cw

[NewGraphEnvironment]

If done: observation validation reports the right miss reasons for groups a bundle
puts on discharge. If never: a mad group's "missed for width" counts are computed
against channel width the classification never used. Nothing is affected until a
bundle actually sets a group to mad, and none does today.

Since #286, lnk_pipeline_classify() passes the bundle's
parameters_habitat_method.csv to fresh, and a mad group classifies on mad_m3s.
lnk_habitat_validate() still builds its predicates with
fresh::frs_habitat_predicates(spp), which is the cw model
(R/lnk_habitat_validate.R:823). It then relaxes width by rewriting s.channel_width
(.lnk_hv_relax(), :772-786), with minimums taken from
ranges$<stage>$channel_width (.lnk_hv_stage_min(), :763). On a mad group:

• the predicate is not the one that classified the segment, so pred_* can disagree
  with the persisted spawning/rearing;
• width relaxation rewrites a column the mad predicate does not reference;
• width_null (.lnk_hv_miss_reason()) tests channel_width, where it should test
  mad_m3s. The persist does not carry mad_m3s (decided in Thread fresh params_method (per-WSG cw/mad) through lnk_pipeline_classify #286), so the validator
  would join it from whse_basemapping.fwa_stream_networks_discharge on
  linear_feature_id.
• The stream-order rearing bypass is skipped for mad groups (Thread fresh params_method (per-WSG cw/mad) through lnk_pipeline_classify #286), so a miss
  reason must not credit it there.

fresh >= 0.35.0 has frs_habitat_predicates(model = "mad"). Resolve each scored
group's model from the bundle's method table, build the predicates per model, and relax
s.mad_m3s for mad groups. test-lnk_habitat_validate.R "the predicate call stays on
the channel-width model" pins today's behaviour and should change with it.

As landed (#301)

• Each scored group's model comes from the cfg method table, resolved by fresh's .frs_habitat_models(), now shared with classify. mad_m3s is joined from the discharge table on linear_feature_id, as proposed.
• Labels kept (decision at the plan gate): fails_width / width_null mean the group's size dimension. observations gains model and mad_m3s, and summary gains model, so a reader can split them.
• New reason no_mad_threshold (not in the original proposal): a mad miss that a MAD range alone would admit, for BT, GR, KO and RB, which have none. Tuning and adding those thresholds is Tune MAD (discharge) thresholds for BT, GR, KO and RB rather than leave them missing #302.
• The stream-order bypass needed no change: the validator never credits it on either model.
• Live check on ADMS: cw output is unchanged; on mad, main had labelled 27 BT rearing misses post_predicate (data-raw/logs/habitat_validate_299/).