Fold subsurfaceflow into natural barriers for per-species obs/habitat lift

[NewGraphEnvironment]

Summary

Subsurfaceflow positions (FWA edge_type IN (1410, 1425)) were unioned into fresh.streams_breaks with label blocked but never entered <schema>.natural_barriers. lnk_barrier_overrides() runs on natural_barriers, so the per-species observation/habitat-upstream lift never fired for subsurfaceflow — every position hard-blocked all species, even where bcfishpass lifts it.

Folds the standalone .lnk_pipeline_prep_subsurfaceflow body into .lnk_pipeline_prep_natural, gated on cfg$pipeline$break_order containing "subsurfaceflow". bcfishpass's source-of-truth (model_access_bt.sql, model_access_ch_cm_co_pk_sk.sql) defines the natural-barrier set as gradient + falls + subsurfaceflow + user_definite. prep_natural already implemented three of those four — the missing entry was the bug.

Six prep helpers → five.

Behaviour

• bcfishpass bundle (subsurfaceflow opted in): subsurfaceflow positions appended to natural_barriers → per-species lift applies → bcfp parity restored on Lower-Fraser/coastal WSGs.
• default bundle (subsurfaceflow not in break_order): early-return inside prep_natural skips the entire subsurfaceflow code path. No <schema>.barriers_subsurfaceflow table, no natural_barriers rows, no streams_breaks rows downstream. Verified bit-identical default rollup (0 of 581 rows changed).

Verification (15-WSG tar_make)

WSG/sp	metric	pre	post
HARR CH	rearing_stream km	-14.8%	-0.32%
HARR CO	rearing_stream km	-13.3%	-0.31%
HARR ST	rearing_stream km	-11.6%	-0.24%
HARR BT	rearing_stream km	-10.4%	-4.19%
LFRA CH/CO/ST	all	gap	<0.6%
HARR blkey 356286055	BT km	0.0	6.509
HORS BT	rearing_stream km	-7.68%	-7.68% (separate mechanism — follow-up)

Distribution: 162 → 171 within ±5%, 123 within ±1%. Per-WSG n_over_5: HARR 11→4, LFRA 6→4, HORS unchanged.

Reproducibility: Second tar_make byte-identical to first. 0 of 1057 link_value rows differ. digest::digest(link_value) matches across runs (5a641892b82604259b0ba168ea093661).

Test plan

• devtools::test(filter = "lnk_pipeline_prepare") — 44/44 pass
• /code-check — 3 rounds, fragile cross-statement regex tightened to per-statement match
• HARR single-WSG pre-flight — 89.5 s, blkey 356286055 → 6.509 km BT
• Full 15-WSG tar_make — 53m 2s, 33/33 targets
• Default-bundle invariance — 0 of 581 rows changed
• Reproducibility re-run — 52m 55s, byte-identical

Architecture note

.lnk_pipeline_prep_subsurfaceflow (added in #82) treated subsurfaceflow as a parallel pipeline phase needing its own helper. prep_natural was always the right home — bcfp's natural-barrier union is gradient + falls + subsurfaceflow, and prep_natural already built the first two. Folding back keeps a single function responsible for "natural barriers in bcfp's sense."

Follow-ups (out of scope)

• HORS rearing_stream gap (~7%) is unchanged by this PR — separate mechanism, will diagnose in a follow-up issue.
• Species-dependent dam-class handling in dimensions.csv #83 (anthropogenic dam design) — unaffected by this PR; barrier/potential labels for crossings still don't gate by default.

Fixes #88
Relates to NewGraphEnvironment/sred-2025-2026#24