fix(globi): hasHost with a pollinator subject is forage, not pest_pressure

[vbeefy]

Problem

lib/globi-classify.js::resolveVariable resolves GloBI's hasHost term with a generic rule:

// Invertebrate has host plant = pest pressure
if (srcBio === 'invertebrate' && tgtIsPlant) → pest_pressure

A bee is an invertebrate, and its "host plant" is the plant it forages on. So every bee→plant hasHost record became a harmful crop-pest claim — a sign inversion: a mutualist recorded as harm.

This is the most damaging error class for an agroecology knowledge base. It doesn't just add noise — it inverts the recommendation ("this bee damages your dandelion" instead of "this bee pollinates it"). It also corrupts anything downstream that treats "organism harming a crop" as a pest node.

Fix

A pollinator guard placed before the generic invertebrate rule:

• subject primary_role='pollinator' OR a POLLINATOR_FAMILIES bee family (Apidae, Halictidae, Andrenidae, Megachilidae, Colletidae, Melittidae, Stenotritidae)
• with a plant object ⇒ pollination / beneficial

Both signals are needed. The family fallback matters because much of the corpus carries primary_role='unclassified'; the role check independently catches non-bee flower-visitors such as Syrphidae (hoverflies), which no bee-family list would.

Not over-correcting

Regression tests pin the paths that must not change:

• a genuine herbivore → plant still ⇒ pest_pressure
• a fungal pathogen → plant still ⇒ pathogen_pressure

Retroactive re-derive

fix-globi-pollinator-category.cjs corrects already-loaded claims by re-running the fixed classifyTriple over the affected candidate set and writing back only where the derived category changes.

Two deliberate properties:

• Self-limiting — a real pest re-derives to pest_pressure and is left untouched.
• One source of truth — it reuses the classifier instead of duplicating the mapping, so it can't drift from it.

Dry-run by default; every change is revision-logged (fully reversible).

Same bug class as hasVector

GloBI's broad terms (hasHost, hasVector) get resolved by generic bio_category rules, and each such rule can silently mis-brand mutualists — cf. the earlier "hasVector is dispersal, not disease" correction, which had mislabeled seed-dispersers as disease vectors. Useful heuristic when adding or auditing one of these rules: ask which mutualist it would mis-brand.

Test plan

node --test backend/lib/globi-classify.test.js → 10/10 pass (verified against this branch's base).

🤖 Generated with Claude Code