MAD (discharge) ranges for BT, GR, KO and RB in default_tuned, scored on held-out groups (#302)

CLAUDE.md

@@ -11,6 +11,31 @@ Experimental package — breaking all the time and loving the learning curve. St
 **Prefix:** `lnk_`
 **Branch:** `main` (current version: `DESCRIPTION` / [`NEWS.md`](NEWS.md))
 
+## Status (2026-10-02, late) — MAD ranges for BT, GR, KO and RB in `default_tuned` (#302)
+
+**`default_tuned` now carries mean-annual-discharge ranges for the four species
+`default` leaves without one; they are inert while every group is on `cw`.**
+- Values (m³/s, maxima open): BT 0.078 / 0.078, GR 0.96 spawn / 0.97 rear, RB 0.011 /
+  0.019, KO 0.57 (unscored). Scored on held-out WSGs by a rule fixed before the build;
+  every P10 → P05 loosening was refused except RB spawning.
+- **Cost before moving any group to `mad`:** at least −33 % BT and −82 % GR stream
+  rearing against `cw` on the held-out WSGs. Stream size and sampling effort are not
+  separated; a size-stratified core is the next tuning. `research/habitat_thresholds.md`,
+  "MAD (discharge)"; archive `planning/archive/2026-10-issue-302-mad-thresholds/`.
+
+**Facts not worth re-deriving:**
+- **The variants harness scores MAD ladders now:** `model` / `set` columns, an anchor
+  rung (cw base → first `mad` rung, taken by construction), `_min` loosens downward, a
+  MAD ladder's core is its rungs only (the cw base cuts on width), `--floor`,
+  `--working-prefix` derived from `--prefix`. #284's re-score reproduces all ten outputs.
+- **Build classifiers from fresh's compiled SQL, never from `rules.yaml`.** fresh's
+  main rear predicate ignores a W rule's `wetland_ha_min`, and its `L` rules include
+  reservoirs (fresh issue drafted, not filed). A rear size range gates the
+  `lake_rearing` / `wetland_rearing` bucket columns, never `rearing`.
+- **`elevation_adjusted.csv`'s band classes are not the bands.** A segment takes the first
+  step whose flags differ, so an anchor's `removed` class is only what no rung restores.
+  Quote `bands_pooled.csv` for a band's rate.
+
 ## Status (2026-10-02) — the validator scores each group on its own model (#299)
 
 **`lnk_habitat_validate()` re-tests a `mad` group on discharge, not width.**

RUNBOOK.md

@@ -702,12 +702,22 @@ next dispatch.**
   group. `log_input` fingerprints it.
 - **What `mad` changes** (fresh's rules, after bcfishpass): stream rules that
   inherit thresholds test `mad_m3s` against `*_mad_min/max` instead of channel
-  width, so species with no MAD thresholds (BT, GR, KO, RB) lose all stream
-  habitat; the rule-level `channel_width` river-polygon bypass is ignored; SK/KO
+  width, so species with no MAD thresholds (BT, GR, KO and RB in `default` and
+  `bcfishpass`) lose all stream habitat; the rule-level `channel_width` river-polygon bypass is ignored; SK/KO
   lake rearing is polygon membership. **Lake, wetland and `thresholds: false`
   rules inherit nothing under either model**, so BT keeps its wetland and
   1050/1150-edge rearing in a `mad` group (ADMS: 63.5 km, all inside
-  waterbodies). link also skips the stream-order rearing bypass
+  waterbodies). That holds for the `rearing` flag. The separate `lake_rearing` /
+  `wetland_rearing` bucket columns (fresh's `build_wb_pred()`) are gated by the
+  species' rear size range whenever it has one: by width under `cw`, and by
+  discharge under `mad`. A species with no rear MAD range keeps them on polygon
+  membership alone. fresh's main predicate also ignores a W rule's
+  `wetland_ha_min` (only `build_wb_pred()` applies it).
+- **`default_tuned` carries MAD ranges for BT, GR, KO and RB** (#302); `default`
+  and `bcfishpass` do not. They are inert until a group is moved to `mad`. At
+  the landed values a `mad` group keeps less stream rearing than `cw` (held-out:
+  at least −33 % BT and −82 % GR, −4 % RB). Read `research/habitat_thresholds.md`
+  before moving one. link also skips the stream-order rearing bypass
   (`frs_order_child`) for a `mad` group — bcfp applies it in its cw branch only.
   fresh does not implement bcfp's `stream_order >= 8` spawning bypass.
 - **Discharge coverage is uneven.** A segment with NULL `mad_m3s` fails every mad

data-raw/habitat_score/README.md

@@ -1,6 +1,7 @@
 # habitat_score
 
-The inputs for scoring habitat-threshold variants against fish observations (#284 step 5).
+The inputs for scoring habitat-threshold variants against fish observations (#284 step 5,
+and the MAD ranges of #302).
 The rule and the reasoning behind it are in `research/habitat_thresholds.md`, under
 "Scoring design". Two scripts use these files:
 
@@ -24,6 +25,31 @@ row changes exactly one cell of `default`'s `parameters_habitat_thresholds.csv`.
 | `obs_stage` | which observations count in the band: `any`, `spawn` or `rear` (the stage the threshold was calibrated on) |
 | `step_from` | the neighbour nearer `default` in the ladder; the band is the difference between the two |
 | `equals_bundle` | a shipped bundle whose thresholds this variant must equal byte for byte (checked at build) |
+| `model` | optional (#302): `cw` (empty) or `mad`. A `mad` variant's bundle also carries a `parameters_habitat_method.csv` with every WSG holding a role for its species on `mad` |
+| `set` | optional (#302): further cells the variant fixes, `col=value;col=value`, held constant along a ladder (a MAD rung's open maximum; the spawning range a clustered rearing ladder needs) |
+
+### MAD ladders (#302)
+
+A species with no MAD range has no "current value" to walk out from: under `mad` it has no
+stream habitat at all. So a MAD ladder starts with an **anchor**, a rung that steps from the
+`cw` base straight to `mad` at the tightest candidate. The score marks it
+`take (anchor: model change)`, reports its bands both ways (what `mad` keeps that `cw` drops,
+and the reverse) and decides nothing with them. The walk then scores the loosening rungs
+beyond it as #284 did. The core is the habitat every ranged rung keeps: the `cw` base is
+left out of it, because it tests a width floor (and fails NULL widths) that no `mad` rung
+tests, so a core cut by it would differ from the bands on an axis no step is about.
+
+A `mad` variant's `set` may name only `*_mad_min` / `*_mad_max` cells, so the rungs carry
+`default`'s gradient cutoffs and every band differs from its neighbour in discharge alone. (`default_tuned`'s BT
+`rear_gradient_max` 0.1349 was scored on its own, in #284.) The score checks that every
+`set` cell holds its value, that a ladder holds `set` fixed past its anchor, that each step
+reads the same species, column, flag and stage as its `step_from`, and that each bundle's
+method table puts its WSGs on the variant's `model`.
+
+A second scoring run gets its own `--prefix`; its working networks are then
+`working_<prefix><wsg>` (#284's, under `score284_`, stay `working_score_<wsg>`).
+
+The inputs for #302 are `variants_302.csv` and `wsg_roles_302.csv`.
 
 ## `wsg_roles.csv`
 

data-raw/habitat_score/variants_302.csv

@@ -0,0 +1,20 @@
+variant,species_code,column,value,flag,obs_stage,step_from,equals_bundle,model,set
+default,,,,,,,default,cw,
+bt_rear_mad_p10,BT,rear_mad_min,0.078,rearing,any,default,,mad,rear_mad_max=9999;spawn_mad_min=0.027;spawn_mad_max=9999
+bt_rear_mad_p05,BT,rear_mad_min,0.027,rearing,any,bt_rear_mad_p10,,mad,rear_mad_max=9999;spawn_mad_min=0.027;spawn_mad_max=9999
+bt_rear_mad_p02,BT,rear_mad_min,0.0085,rearing,any,bt_rear_mad_p05,,mad,rear_mad_max=9999;spawn_mad_min=0.027;spawn_mad_max=9999
+bt_spawn_mad_p10,BT,spawn_mad_min,0.078,spawning,any,default,,mad,spawn_mad_max=9999;rear_mad_min=0.027;rear_mad_max=9999
+bt_spawn_mad_p05,BT,spawn_mad_min,0.027,spawning,any,bt_spawn_mad_p10,,mad,spawn_mad_max=9999;rear_mad_min=0.027;rear_mad_max=9999
+bt_spawn_mad_p02,BT,spawn_mad_min,0.0085,spawning,any,bt_spawn_mad_p05,,mad,spawn_mad_max=9999;rear_mad_min=0.027;rear_mad_max=9999
+gr_rear_mad_p10,GR,rear_mad_min,0.97,rearing,any,default,,mad,rear_mad_max=9999;spawn_mad_min=0.095;spawn_mad_max=9999
+gr_rear_mad_p05,GR,rear_mad_min,0.095,rearing,any,gr_rear_mad_p10,,mad,rear_mad_max=9999;spawn_mad_min=0.095;spawn_mad_max=9999
+gr_rear_mad_p02,GR,rear_mad_min,0.0026,rearing,any,gr_rear_mad_p05,,mad,rear_mad_max=9999;spawn_mad_min=0.095;spawn_mad_max=9999
+gr_spawn_mad_p10,GR,spawn_mad_min,0.96,spawning,any,default,,mad,spawn_mad_max=9999;rear_mad_min=0.095;rear_mad_max=9999
+gr_spawn_mad_p05,GR,spawn_mad_min,0.095,spawning,any,gr_spawn_mad_p10,,mad,spawn_mad_max=9999;rear_mad_min=0.095;rear_mad_max=9999
+gr_spawn_mad_p02,GR,spawn_mad_min,0.0025,spawning,any,gr_spawn_mad_p05,,mad,spawn_mad_max=9999;rear_mad_min=0.095;rear_mad_max=9999
+rb_rear_mad_p10,RB,rear_mad_min,0.019,rearing,any,default,,mad,rear_mad_max=9999;spawn_mad_min=0.011;spawn_mad_max=9999
+rb_rear_mad_p05,RB,rear_mad_min,0.0094,rearing,any,rb_rear_mad_p10,,mad,rear_mad_max=9999;spawn_mad_min=0.011;spawn_mad_max=9999
+rb_rear_mad_p02,RB,rear_mad_min,0.0048,rearing,any,rb_rear_mad_p05,,mad,rear_mad_max=9999;spawn_mad_min=0.011;spawn_mad_max=9999
+rb_spawn_mad_p10,RB,spawn_mad_min,0.05,spawning,spawn,default,,mad,spawn_mad_max=9999;rear_mad_min=0.0094;rear_mad_max=9999
+rb_spawn_mad_p05,RB,spawn_mad_min,0.011,spawning,spawn,rb_spawn_mad_p10,,mad,spawn_mad_max=9999;rear_mad_min=0.0094;rear_mad_max=9999
+rb_spawn_mad_p02,RB,spawn_mad_min,0.0055,spawning,spawn,rb_spawn_mad_p05,,mad,spawn_mad_max=9999;rear_mad_min=0.0094;rear_mad_max=9999

data-raw/habitat_score/wsg_roles_302.csv

@@ -0,0 +1,22 @@
+watershed_group_code,species_code,role
+REVL,BT,held_out
+ELKR,BT,held_out
+KOTR,BT,held_out
+UARL,BT,held_out
+MURR,BT,held_out
+UBTN,BT,held_out
+LHAF,BT,held_out
+PARS,BT,in_sample
+KOTL,BT,in_sample
+UBTN,GR,held_out
+MURR,GR,held_out
+LHAF,GR,held_out
+PARS,GR,in_sample
+SIML,RB,held_out
+OKAN,RB,held_out
+KETL,RB,held_out
+ELKR,RB,held_out
+REVL,RB,held_out
+KOTR,RB,held_out
+PARS,RB,in_sample
+KOTL,RB,in_sample