[NewGraphEnvironment]

Summary

• A bundle chooses its discharge source: pipeline: discharge_source: fwapg | wet (default fwapg). The source is threaded through prepare, the run log (<schema>.log.discharge_source), the validator, the export manifest (discharge.source, discharge.wet_load), the scoring harness and the MAD producers. New export lnk_preflight_discharge(), loader data-raw/discharge_wet_load.R, and host prep in study_area_run.sh / cypher_prep.sh. The fwapg SQL is byte-identical to main's.
• default_tuned reads wet; BT spawn_mad_min / rear_mad_min 0.078 → 0.043.
  • Rearing was scored on held-out groups (P10 → P05 refused).
  • Spawning takes the same any-stage percentile, because the held-out core holds only 3 spawning-staged locations.
  • default and bcfishpass stay on fwapg.
  • No habitat flag moves while every group is on cw: on NATR, 0 of 189,200 rows changed.
• Why the values moved: coverage, not the source. On BT locations both sources value, P10 is 0.075 on wet and 0.076 on fwapg. wet adds nine Skeena groups fwapg carries no discharge for, and their water is smaller.
• At the shipped values mad is the closer model for BT.
  • On held-out groups, the water only cw keeps is used at 0.18 (rearing) and 0.14 (spawning) of the size-matched rate. The water only mad keeps is used at 1.64 and 0.69.
  • mad keeps 19.8 % less BT rearing than cw (Main stems with no discharge drop out of mad habitat #305: 32 % at 0.078 on fwapg).
• Write-up: research/habitat_thresholds.md, "MAD on wet's discharge (Read discharge from wet in classify and the habitat export, and re-tune BT's MAD range on it #331)". Release 0.66.0.

Related Issues

• Closes Read discharge from wet in classify and the habitat export, and re-tune BT's MAD range on it #331
• Follow-ups: Re-tune default_tuned's GR, RB and KO MAD minima on wet's discharge #343 (GR, KO and RB on wet), wet_db_load(): record which parts were loaded (key, md5) in the database wet#88 (per-part load record)
• Relates to NewGraphEnvironment/sred#24

Test plan

• Full devtools::test() at 5bd9b325: 2962 pass, 7 fail.
  • 2 failures were this branch's (deefda33 fixes them).
  • 3 need the bcfishpass tunnel on :63333, which was down: test-lnk_db_conn.R:10, test-lnk_wsg_resolve.R:143,154.
  • 2 are the golden dams pin (test-lnk_pipeline_prepare.R:958-959). Local cabd.dams went from 2,594 to 2,592 rows upstream, and this branch doesn't touch the dams code. Re-pin deliberately.
• devtools::check(--no-tests): 0 errors. Its 3 warnings and 4 notes are all on main (non-ASCII in untouched files, Rd links, setNames imports).
• Scoring builds on one segmentation:
  • Build A: BT ladders on wet and fwapg, base default.
  • Build B: cw against mad on default_tuned.
  • Both scored with Tune MAD (discharge) thresholds for BT, GR, KO and RB rather than leave them missing #302's floor of record (--floor=expected).
  • Logs: data-raw/logs/habitat_score_331*.
• cw unchanged on NATR: one prepare, mad_m3s re-joined from each source, 0 habitat rows differ (data-raw/logs/discharge_wet_331/cw_unchanged_natr.csv).
• /code-check on the branch, five rounds. Rounds 1–3 covered the code and ended by enumeration. Rounds 4–5 covered the Phase 5 values and the write-up: no code defects, and published claims were narrowed to their evidence, ending with a grep of each claim class.

Notes

• A host that runs default_tuned needs wet's table: LNK_LOAD=loadall Rscript data-raw/discharge_wet_load.R.
  • That script loads the annual pick only.
  • On a host holding wet's full load, --force or a differing --expect-load replaces it with the pick.
• default_tuned's config_hash changes.
• While Build B ran, the local DB volume filled. 62 stale working_<wsg> / zz311_* schemas were dropped; every score* schema was kept.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WJE2szSCkeV9R3qKhuSwyx