Score each watershed group on its own habitat model in lnk_habitat_validate() (#299)

[NewGraphEnvironment]

Summary

• lnk_habitat_validate() re-tests each observation's segment on the habitat model its watershed group classified on (cw or mad), from the method table in the cfg it is given. The model is resolved by fresh's own .frs_habitat_models(), which lnk_pipeline_classify() now shares.
• On a mad group, the predicates come from frs_habitat_predicates(model = "mad"), size relaxation moves mad_m3s (joined from fwa_stream_networks_discharge on linear_feature_id, since the persist does not carry it), and width_null tests discharge. Reason labels are shared across models; observations gains model and mad_m3s, summary gains model.
• New reason no_mad_threshold: a mad miss that a MAD range alone would admit, for the species that have none (BT, GR, KO, RB). A miss whose gradient also fails reads fails_gradient_and_width; one with no discharge on its line reads width_null, as on cw.
• Preflight now also asserts frs_habitat_predicates(model) and .frs_habitat_models. The miss-binning driver carries model.

Related Issues

• Fixes lnk_habitat_validate() scores mad watershed groups as if they were cw #299
• Unblocks Score the discharge (mad) habitat model against observations, as #284 did for gradient #300 (scoring the mad model against observations) and Tune MAD (discharge) thresholds for BT, GR, KO and RB rather than leave them missing #302 (adding MAD thresholds)
• Relates to NewGraphEnvironment/sred#24

Test plan

• Full suite: 2,394 tests, 0 failed, 0 skipped (16 warnings, the existing baseline)
• R CMD check: 0 errors. The 3 warnings and 2 notes are pre-existing (non-ASCII in other files, Rd cross-refs, setNames/getFromNamespace globals) and touch no line this branch adds
• cw unchanged: ADMS on fresh_default, branch vs main, observations identical apart from the new columns and summary identical apart from model (94 locations)
• mad live: ADMS modelled on mad into a scratch schema, then validated with branch and with main. Main calls 27 BT rearing misses post_predicate, the branch no_mad_threshold. See data-raw/logs/habitat_validate_299/README.md

Notes

• Review: one plan review and four code-check rounds (five agents). Two defects sat inside the no_mad_threshold fixes: it relaxed gradient too, and it ignored a NULL discharge. Round 3 named the mechanism, a hand-copied relaxation ladder, and the loop ended on two enumerations rather than a quiet round. The first covered every miss arm × model × NULL pattern, evaluated in Postgres. The second was a test of the missing-range rule against fresh, case by case.
• The planned acceptance metric did not discriminate. "Persisted TRUE, predicate FALSE" is 0 on main as well. A wrong model shows up as post_predicate, which is now recorded in CLAUDE.md.
• Known limit: a run classified with a different method table (or lnk_pipeline_classify(method_csv =)) is not detected; swap the table on the cfg passed to the validator.
• Follow-up: Tune MAD (discharge) thresholds for BT, GR, KO and RB rather than leave them missing #302 tunes and adds MAD thresholds for BT, GR, KO and RB.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WJE2szSCkeV9R3qKhuSwyx