Tighten NPC interaction validation and consistency

docs/integrations/dnd-npc-interaction-artifacts.md

@@ -33,23 +33,57 @@ array may be empty. Each item has exactly `name`, `kind`, and `source_refs`:
 `name` is the canonical display name from the required NPC registry.
 `source_refs` contains one or more current-source ranges with required
 `source_id`, `start_unit_id`, and `end_unit_id`; unit IDs are positive integers.
+During extraction, every range must be wholly contained in the current accepted
+chunk. This prevents a candidate from citing valid units that were not presented
+to that extraction call.
 Unknown fields are rejected.
 
 ## Interaction Categories
 
 `kind` is exactly one of:
 
-- `mentioned`: the named NPC is referenced without stronger participation.
-- `noncombat_presence`: the NPC is present in the current scene without a
-  dialogue or combat classification.
-- `dialogue`: the NPC participates in spoken interaction.
-- `combat_ally`: the NPC participates in combat aligned with the party.
-- `combat_opponent`: the NPC participates in combat against the party.
-- `other`: a transcript-supported interaction outside the bounded categories.
+| Kind | Meaning |
+| --- | --- |
+| `mentioned` | The NPC is referred to, but is not established as present or communicating in the evidenced passage. |
+| `noncombat_presence` | The NPC is present and relevant to the passage but does not meaningfully participate in dialogue or combat. |
+| `dialogue` | The NPC speaks, responds, or is directly engaged in a meaningful non-combat exchange. |
+| `combat_ally` | The NPC actively participates in combat on the party's side. |
+| `combat_opponent` | The NPC actively participates in combat against the party. |
+| `other` | The transcript clearly establishes a direct NPC occurrence that fits none of the preceding kinds. |
+
+`other` is a residual category for positively evidenced activity, not a fallback
+for uncertain classification. When activities overlap, active combat
+participation outranks dialogue, presence, and mention; dialogue outranks
+non-combat presence and mention; and non-combat presence outranks mention.
+Combat alignment is not resolved by precedence: a meaningful change between
+ally and opponent creates separate occurrences.
 
 These categories do not encode summaries, relationships, state, motives, or
 unobserved events.
 
+## Occurrence Boundaries And Ordering
+
+One occurrence represents one NPC, one kind, and one locally coherent passage
+within one accepted chunk. Repeated evidence belongs to the same occurrence
+only while it supports the same uninterrupted activity. A kind change, combat
+alignment change, intervening scene or meaningful absence, or transition from
+mention to presence starts a new occurrence. Occurrences never span chunks, and
+merge or normalization never semantically combines nearby, overlapping, or
+cross-chunk records.
+
+Normalization orders records by:
+
+1. earliest valid source-document position;
+2. the NPC identity comparison key;
+3. the exact canonical NPC display name;
+4. interaction kind in lexical order; and
+5. the complete canonical source-reference sequence, ordered by source ID and
+   the source-document positions of each range's start and end.
+
+Only records with identical canonical names, kinds, and complete valid evidence
+sequences are duplicates. Different categories, ranges, or separately grounded
+occurrences remain separate.
+
 ## Evidence, Registry, And Normalization
 
 The registry proves only the canonical NPC identity. Its source references are
@@ -59,11 +93,8 @@ classification.
 
 The extractor receives a names-only registry projection such as
 `{"npcs":[{"name":"Mira Thorn"}]}`. The normalizer uses the full immutable
-registry for exact canonical-name lookup. It orders source references,
-stable-sorts occurrences by their earliest source-document position, and
-collapses only exact duplicates with the same canonical name, kind, and complete
-valid evidence. Different categories, distinct ranges, and separately grounded
-occurrences remain separate; no semantic merge is performed.
+registry for exact canonical-name lookup. It canonicalizes source references
+and applies the ordering and exact-duplicate rules above.
 
 ## Production Pipeline
 
@@ -102,8 +133,10 @@ copying registry names, source ranges, or payload content into the manifest.
 
 The default extract chain is `generic/valid_json`, interaction shape, registry,
 and source-reference validation, `generic/valid_json_schema`, then warning-only
-source relatedness. The normalize chain adds normalized invariants after schema
-validation and before relatedness. The codec metadata contains only
+source relatedness. The normalize chain runs normalized invariants after
+registry validation and before source-reference and schema validation, followed
+by relatedness. Normalizer and relatedness warnings are bounded and end with an
+omission summary when necessary. The codec metadata contains only
 `interaction_count`. Extractor metadata identifies its prompt and private
 response schema; component-local checkpoint identities include the names-only
 registry projection where relevant. Generated registry identity stays in