#284 step 5: score BT rear_gradient_max variants on held-out WSGs (0.1349 holds)

[NewGraphEnvironment]

Summary

• Research: calibrate CH and BT gradient and channel-width thresholds from observations #284 step 5: BT rear_gradient_max 0.1349 in default_tuned is scored and held.
  • It was scored on eight watershed groups held out from its calibration.
  • The steps to 0.1349 add rearing that BT use at 0.60 and 0.71 of the core rate (0.75 and 0.89 at the same elevation).
  • The step to 0.1449 does not hold up: 8 found where 25 were expected.
  • default_tuned is unchanged. Results are in research/habitat_thresholds.md, "Step 5".
  • What that means for habitat: 0.1349 adds 1,105 km of BT rearing (+5.8 %) across the 12 watershed groups, from 2 % (BABL) to 9 % (KOTL, UARL). The per-group table is habitat_change.csv.
• New export lnk_habitat_validate_band(). It measures observation density on the segments a threshold step moves, against the core every step agrees on. It joins on the full key and stops if segmentation differs or a schema holds no habitat for a WSG.
• Harness, all inputs as data (data-raw/habitat_score/):
  • data-raw/habitat_variants_build.R models a drainage closure once, then re-classifies each threshold variant on that shared segmentation. Resume only reuses clean, same-commit bases, and built.csv records which bundle each schema came from.
  • data-raw/habitat_variants_score.R validates each variant and applies the fixed rule. It also writes a taper, an elevation split, an elevation-adjusted ratio, the bridge split and an expected-count floor beside the rule.
• FISS absence taxa are now data (data-raw/fiss_absence_taxa.csv), and the absence and pooling loaders are shared (habitat_validate_inputs.R). The Validate modelled habitat against fish observations per species and watershed group #283 outputs are byte-identical, with knowledge pinned at 508bf44.

Related Issues

• Closes Research: calibrate CH and BT gradient and channel-width thresholds from observations #284
• Spun out NewGraphEnvironment/knowledge#28 (habitat weights instead of cutoffs, biology first)
• Relates to NewGraphEnvironment/sred#24

Test plan

• devtools::test(): 0 fail, 2192 pass, the same 16 warnings as the baseline
• devtools::check(): 0 errors. The 3 warnings and 2 notes are pre-existing (non-ASCII in older R files, missing Rd links, undocumented presence); none touch this diff.
• lintr: no lints in the new files
• Band guards mutation-tested: segmentation guard removed → red; bare id_segment join → red; habitat-presence guard removed → red
• BULL pre-flight, then the full build (run 20260930_014921-53610765, 139.6 min), with post-conditions verified in the DB:
  • 23 base WSGs, all clean at a11a001;
  • 12 focal WSGs per variant;
  • default re-classify digests equal to the base run.
• Bands reconcile with the rollup: default → 0.1349 adds 585 km, equal to 422.2 + 163.0 km of bands.
• /code-check: 5 rounds plus an enumeration (planning/archive/2026-09-issue-284-step5-scoring/review-*.md)

Notes

• The design changed before any band was seen. A counts-only power check showed the planned held-out set could decide one step in six. So:

  • the BT held-out set was widened;
  • CH was dropped, and its verdicts stay "keep", unscored;
  • an underpowered step changes nothing.

  Operator's calls, recorded in the research doc with the counts.

• The n ≥ 10 floor is open. Counting locations found cannot refuse a band fish avoid; counting locations expected at the core rate can. Both readings are in verdict.csv. They agree on this run, and the choice stays open for the next tuning.

• Elevation confounds gradient. Core BT rearing thins from 24.6 to 10.2 locations per 100 km between the low and high thirds of each WSG's rearing, and the steep bands sit high.

• Clustering moves a little habitat the other way. Loosening the cutoff removed 1.1 km of rearing (against 1,365 km added). The nesting guard tolerates up to 1 % and reports the km.

• The score284_* schemas are kept in local docker fwapg until this merges. Dropping them is a separate step.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WJE2szSCkeV9R3qKhuSwyx