Reorder shared D&D extraction prompt messages

docs/internal/llm.md

@@ -98,6 +98,18 @@ D&D prompt content. `internal/framework/promptfs` provides the domain-neutral
 filesystem composition helper used to combine module-owned files with shared
 domain prompt fragments.
 
+The D&D spell, NPC, and combat-turn extractors each declare an ordered prompt
+asset manifest. The manifest lists the package-owned YAML and Markdown files,
+then the exact shared fragments rendered by that prompt; the same ordered list
+drives both filesystem mounting and the prompt fingerprint. Shared extraction
+evidence and in-world identity messages precede the cacheable transcript and
+campaign-reference messages. Spell and combat prompts then render immediate
+resolution and the cacheable NPC registry before their lane-specific material;
+the NPC prompt renders its task and instructions instead. The identity,
+transcript, and campaign-reference messages are ephemeral in all three prompts,
+and the NPC registry is ephemeral where it is used by spell and combat prompts.
+Unused shared assets are neither mounted nor fingerprinted.
+
 Schema helpers load embedded JSON Schema with identity and digest metadata,
 return defensive copies, and expose a diagnostics map that omits schema bytes.
 The small framework registry contains only generic test schemas; production
@@ -113,8 +125,8 @@ input and an optional `application/json` `npcs` input. The extractor generates
 the catalog input from its prepared
 effective catalog as `{"spell_names":[...]}` using sorted canonical names only.
 The shared D&D prompt assets include a generic NPC grounding fragment directly
-after the campaign reference message for both extractors. When an NPC registry
-is bound, the
+after the campaign reference message for spell and combat prompts. When an NPC
+registry is bound, the
 domain registry boundary strictly decodes and identity-validates one durable
 artifact, re-encodes canonical JSON, and generates a semantic digest over
 those bytes. The unbound input is exactly `{"npcs":[]}`. Input digests cover

docs/internal/modules.md

@@ -47,9 +47,16 @@ LLM-backed extensions own their prompt definitions and response schemas under
 package-local embedded assets. Shared filesystem composition belongs in
 `internal/framework/promptfs`; reusable D&D prompt fragments, reference
 declarations, prompt-input assembly, and source-unit helpers belong in
-`internal/modules/dnd/shared`, which also owns bounded D&D diagnostics. Stage
-contracts expose only Notarius structured-
-completion types, not Scriptorium public types.
+`internal/modules/dnd/shared`, which also owns bounded D&D diagnostics. The
+spell, NPC, and combat-turn extractors use ordered package-local prompt
+manifests for both rendering and prompt fingerprinting, so only the shared
+fragments each prompt actually renders participate in either operation. Their
+common prompt order is system, extraction evidence, in-world identity,
+transcript, and campaign references, with ephemeral cache control on identity,
+transcript, and references. Spell and combat prompts append immediate
+resolution and an ephemeral NPC registry before lane-specific material; the
+NPC prompt appends its task and instructions. Stage contracts expose only
+Notarius structured-completion types, not Scriptorium public types.
 
 Reference material may inform a module or prompt but must not become source
 evidence. The resolver and materializer behavior is described in
@@ -182,15 +189,14 @@ source evidence.
 
 The prompt limits each cast to its declaration and immediate resolution; it
 does not follow summoned creatures, persistent effects, or other downstream
-consequences through the scene. Source references must collectively support all
-reported facts, using separate narrow ranges when immediate evidence is
-non-contiguous. The prompt grounds caster names in in-world identities, using
-the player and party references to disambiguate transcript speakers without
-treating those references as event evidence. Effects describe the session as
-played: model rules knowledge cannot supplement or correct the transcript, and
-nonstandard adjudication is attributed to the GM or table rather than stated as
-a universal rule. Structural source validation remains deterministic; semantic
-claim completeness is enforced through extraction policy and evaluation.
+consequences through the scene. Shared extraction-evidence and identity rules
+require transcript-supported factual claims and the most specific in-world
+caster identity, while campaign references only disambiguate source text.
+Effects describe the session as played: model rules knowledge cannot supplement
+or correct the transcript, and nonstandard adjudication is attributed to the GM
+or table rather than stated as a universal rule. Structural source validation
+remains deterministic; semantic claim completeness is enforced through
+extraction policy and evaluation.
 
 Both the extractor and deterministic catalog validator expose
 the effective base-plus-overlay semantic digest as scoped prepared-component
@@ -223,12 +229,16 @@ The NPC extractor maps private model output to the canonical `dnd.NPCList`,
 assigns source identity and deterministic NPC IDs, and preserves source
 references for deterministic validation. It uses the shared campaign
 references only for disambiguation and does not consume the optional NPC
-registry slot. Its prompt and private response schema are package-owned.
+registry slot. Its prompt and private response schema are package-owned. The
+prompt uses the common evidence, identity, transcript, and campaign-reference
+messages, then the NPC-specific task and instructions; only the identity,
+transcript, and campaign-reference messages carry ephemeral cache control.
 
 ### `internal/modules/dnd/extract/combatturns`
 
 The combat extractor prepares one structured request per supplied chunk using
-the shared transcript, campaign-reference, and NPC-grounding prompt inputs. It
+the shared extraction-evidence, identity, transcript, campaign-reference,
+immediate-resolution, and NPC-grounding prompt inputs. It
 maps the private response to `dnd.CombatTurnList`, assigns the current source
 identity, removes exact duplicate source ranges, and orders turns by valid
 source-document position while preserving malformed candidate fields for
@@ -236,8 +246,11 @@ deterministic validators. Its package-owned private response schema enforces
 only the structural JSON envelope; semantic artifact constraints remain with
 the validator chain. Its prepared metadata and checkpoint fingerprints contain
 only prompt/schema/mapping identities plus an optional NPC registry digest.
-The package exposes typed registration and is included in the production D&D
-registrar with the default combat extraction chain.
+The prompt renders immediate resolution and the NPC registry before the
+combat-specific task and instructions; identity, transcript, campaign
+references, and the NPC registry use ephemeral cache control. The package
+exposes typed registration and is included in the production D&D registrar with
+the default combat extraction chain.
 
 The combat normalizer accepts only the optional structured NPC registry.
 Campaign references remain extractor-only LLM context and are not materialized