Tune MAD (discharge) thresholds for BT, GR, KO and RB rather than leave them missing

[NewGraphEnvironment]

If done: a mad watershed group keeps stream habitat for every modelled species, on
discharge windows set from evidence. If never: any group moved to mad loses all
stream spawning and rearing for BT, GR, KO and RB by construction. That is not a finding
about the fish. Only waterbody rules (wetlands, lakes, the 1050/1150 edges) keep them.

Problem

parameters_habitat_thresholds.csv carries spawn_mad_* for CH, CM, CO, PK, SK, ST and
WCT, and rear_mad_* for CH, CO, ST and WCT only (default, default_tuned and
bcfishpass alike). The gap was inherited from bcfishpass example_newgraph. Under
the mad model, fresh writes FALSE in place of the size test for a species with no
range, as bcfishpass does. Since #299, lnk_habitat_validate() names a miss the missing
range alone explains no_mad_threshold. Misses from the same cause also read
width_null (no discharge on the line) or fails_gradient_and_width (gradient fails
too), so that label is a floor on the loss, not a count of it. On ADMS on mad, 27 BT
locations read it against spawning and 8 more read fails_gradient_and_width
(data-raw/logs/habitat_validate_299/). The operator's direction (2026-10-02,
#299 plan gate): tune and add the thresholds, rather than leave them missing in the long
run.

Proposed Solution

• Measure first. Take the distribution of mad_m3s at BT / GR / KO / RB observation
  locations by stage (pooled DV per species_pooling.csv), alongside accessible
  availability. This is the Research: calibrate CH and BT gradient and channel-width thresholds from observations #284 method with discharge in place of width. As built:
  measured on the 46 discharge-covered fresh_default groups (calibration), not on
  Score the discharge (mad) habitat model against observations, as #284 did for gradient #300's held-out groups, so that the score could stay out-of-sample.
• Literature and FISS for each species' discharge range, recorded in
  research/habitat_thresholds.md as one verdict per threshold.
• Land the values in default_tuned (a thin bundle) first, scored out-of-sample with
  lnk_habitat_validate_band(). default stays untouched until a score supports a move.
• Score before/after with capture and cost (share_*, *_km) and
  lnk_habitat_validate_band(), not with no_mad_threshold. Once a range exists, the
  label cannot fire, so its "after" is 0 whatever the value.

Depends on #299. Relates to #284, #286, #300. As built: the mad scoring harness
#300 needed was built here (operator's call at the plan gate), so #300 can reuse it.

Outcome (2026-10-02, PR #303)

• Landed in default_tuned (m³/s, maxima open): BT 0.078 spawn and rear, GR 0.96
  spawn / 0.97 rear, RB 0.011 spawn / 0.019 rear, and KO 0.57 spawn, unscored (no
  held-out group). They are inert while every group is on cw; default is untouched.
• Scoring: a rule fixed before the build (expected-count floor). Every P10 → P05
  loosening was refused on the held-out groups except RB spawning, so the landed values
  are the P10 rungs, not the P05 candidates.
• Cost before moving any group to mad: at least −33 % BT and −82 % GR stream
  rearing against cw on the held-out groups. Stream size and sampling effort are not
  separated; a size-stratified core is the next tuning.
• Where it lives: research/habitat_thresholds.md, "MAD (discharge) ranges for BT,
  GR, KO and RB (Tune MAD (discharge) thresholds for BT, GR, KO and RB rather than leave them missing #302)"; archive planning/archive/2026-10-issue-302-mad-thresholds/.