Detect one node_id carrying several node_types across records (#356)

docs/CURATION_PLAYBOOK.md

@@ -203,6 +203,50 @@ Those numbers came from measuring the graphs by hand. #359 makes it routine —
 `reports/causal_graph_connectivity.tsv`, one row per graph, arriving with
 **#363**. Once it lands, quote that table rather than the finding counts.
 
+### One `node_id` means one thing — across the whole corpus
+
+`audit-graphs` flags `INCONSISTENT_NODE_TYPE` when one `node_id` carries
+different `node_type`s in different records (#356). It is the only
+**cross-record** check here, and that is exactly why the defect survived: read
+alone, neither record is wrong.
+
+Baselined at **294 occurrences across 63 `node_id`s**. The worst is
+`proton_motive_force` — `STATE`×18, `BIOLOGICAL_PROCESS`×13, `CHEMICAL`×2,
+`CAPACITY`×2, for one concept.
+
+**Why it stopped being cosmetic.** #355 minted `powers` (`METPO:2007900`) gated
+to `subject_types = BIOLOGICAL_PROCESS|STATE`. Two byte-identical assertions now
+behave differently purely by subject typing:
+
+```
+physiology/carboxydotrophic.yaml   proton_motive_force (STATE)     -> atp_synthase   grounds
+physiology/phototrophic.yaml       proton_motive_force (CAPACITY)  -> atp_synthase   blocked_by_node_type
+```
+
+Typing decides groundings now. A disagreement is no longer just untidy.
+
+**A hit is not automatically a defect.** `terminal electron acceptor` is
+deliberately both `CHEMICAL` and `MOLECULAR_FUNCTION`; `mappings/node_grounding.tsv`
+carries a row for each, noting that one METPO class covers both senses and the
+`MOLECULAR_FUNCTION` typing "surfaces the role-of interpretation". Same
+two-senses shape as `reduces` (#330/#333) and the `CAPACITY` table above.
+
+So ask **does this id mean one thing?**
+
+| answer | fix |
+|---|---|
+| yes, and one type is right | normalise the outliers to it |
+| yes, but the right type is arguable (`STATE` vs `BIOLOGICAL_PROCESS`) | decide once, record why, normalise |
+| **no — it means two things** | **split into two `node_id`s**, not one type |
+
+The gradient *is* a state; generating and maintaining it *is* a process. If a
+record means the second, it should not be reusing the id for the first.
+
+**Do not repeat #352's mistake.** The test is not "is this type defensible in
+isolation" — it is "is it compatible with what the record and its predicates
+already assert". #352 spent three review rounds learning that on the
+neighbouring `DISPOSITION_MISTYPED` family, and the grounding is what settles it.
+
 ### `enables` needs a process-or-activity object
 
 Separately from the domain rule above, `enables` (`RO:0002327`) has a

history/infrastructure/inconsistent-node-type/2026-08-12T000912Z-claude-code-779670.yaml

@@ -0,0 +1,43 @@
+history_version: 1
+target:
+  kind: infrastructure
+  path: scripts/audit_causal_graphs.py
+  slug: inconsistent-node-type
+session:
+  id: 2026-08-12T000912Z-claude-code-779670
+  timestamp: '2026-08-12T00:09:12Z'
+  actors:
+  - type: ai_agent
+    name: claude-code
+links:
+  issues:
+  - https://github.com/CultureBotAI/TraitMech/issues/356
+events:
+- type: EDIT
+  outcome: changed
+  sections:
+  - causal_graphs
+  summary: Detect one node_id carrying several node_types across records
+  details: 'Every other check in audit_causal_graphs.py is scoped to one graph. This defect
+    is not: two records disagree about what a node IS, and neither is wrong read alone, which
+    is why nothing caught it. Added node_type_index() plus an INCONSISTENT_NODE_TYPE finding,
+    baselined at 294 occurrences across 63 node_ids. The issue understated the scale by about
+    7x: it reports proton_motive_force typed four ways across 9 records; measured, it is 35
+    records (STATE 18, BIOLOGICAL_PROCESS 13, CHEMICAL 2, CAPACITY 2), and 63 node_ids disagree
+    with themselves corpus-wide including membrane_potential at four types. It stopped being
+    cosmetic when 355 minted powers (METPO:2007900) gated to subject_types BIOLOGICAL_PROCESS
+    or STATE: carboxydotrophic.yaml''s proton_motive_force (STATE) grounds while phototrophic.yaml''s
+    (CAPACITY) is blocked_by_node_type, for byte-identical assertions. Reported per occurrence
+    rather than on a presumed-wrong minority, because nothing here knows which type is right
+    - proton_motive_force splits 18 STATE to 13 BIOLOGICAL_PROCESS and the gradient genuinely
+    is a state while generating it is a process, so the majority is an observation and not
+    a verdict; per-occurrence rows also mean a family clears together the moment it is normalised.
+    The detail leads with node_id so _key discriminates by node within a graph, deliberately
+    NOT with the type set, since a family part-way through a burn-down would otherwise re-key
+    on every step and un-suppress rows nobody has reached yet. A test pins that. Documented
+    that not every hit is a defect: terminal electron acceptor is deliberately both CHEMICAL
+    and MOLECULAR_FUNCTION and node_grounding.tsv carries a row for each saying one METPO
+    class covers both senses, the same two-senses shape as reduces (330/333). Where a family
+    really means two things the fix is two node_ids rather than one type, which is written
+    into the playbook as a decision table. Detection only - no data changed, and the burn-down
+    is the next step. 534 tests pass (4 new), ruff clean.'