The corpus types 'capacity to X' two ways: TRAIT in some records, CAPACITY/STATE in others

[realmarcin]

Raised in review of PR #351 (#334 burn-down).

#351 retyped five nodes from STATE/CAPACITY to TRAIT because their own
descriptions describe dispositions — "Capacity of an organism to grow and survive under
high hydrostatic pressure", "Ability to grow when external phosphate is unavailable",
"Detoxification-mediated tolerance of arsenic oxyanions".

The same concept is typed CAPACITY elsewhere and was left alone:

node	record
salt tolerance — "Capacity to grow and survive under elevated salinity"	slightly_halophilic.yaml
low-pH tolerance	ph_delta.yaml
oxygen tolerance	oxygen_preference.yaml
growth at 4 degrees C	psychrotolerant.yaml

They survive as CAPACITY only because their in-edges did not happen to violate the
enables range — an entirely unrelated fact. So the corpus now types one concept two ways,
and which way depends on whether a different check fired on a neighbouring edge.

Why it matters

CausalNodeTypeEnum membership is what every node-type gate keys on
(subject_types/object_types in mappings/predicate_grounding.tsv), and TRAIT nodes are
additionally what audit-graphs uses as reachability anchors. A concept typed inconsistently
is gated inconsistently and anchors inconsistently.

Suggested approach

1. Sweep CAPACITY and STATE nodes for descriptions matching the disposition pattern —
   capacity to, ability to, tolerance of, growth at/under/without.
2. Decide the rule once and write it into docs/CURATION_PLAYBOOK.md, which Retype six mis-typed dispositions, and route six real states (#334) #351 has already
   started ("a node described as a capacity to, an ability to, or a tolerance of is a
   disposition, i.e. a TRAIT").
3. Retype in one pass, grounding each — Retype six mis-typed dispositions, and route six real states (#334) #351's review showed that an ungrounded new TRAIT
   node silently becomes a reachability anchor and makes UNREACHABLE_FROM_TRAIT fall without
   the graph actually becoming more connected.

Worth doing as its own change rather than inside a range-violation burn-down, since the
trigger for finding them should be the typing rule, not which edges happened to be broken.

[realmarcin]

Additional scope from the #351 re-review.

A repeated grounding within one graph is not detected. #351 grounded growth_external_ph_5_5_9 to METPO:1000478, the same CURIE its own record's ph_delta_high_trait node carries. That was deliberate — it makes the duplication legible rather than leaving two unrelated-looking anchors — but scripts/audit_causal_graphs.py has no check for two nodes in one graph sharing a grounding, so the merge only happens if someone remembers.

Worth adding as part of this issue's sweep: flag graphs where two nodes carry the same grounding, since that is the machine-readable signature of exactly the duplication this issue is about. It would have found the ph_delta_high case without anyone reading the descriptions.