scene-inventory-extractor

name: scene-inventory-extractor description: > End-to-end production-prep workflow: extracts comprehensive scene inventories from narrative writing, extracts continuity inventory and reset-critical state before prompt writing, generates all reference images (characters, locations under multiple angles/conditions, props), produces start/end/keyframe shot references with consistency verification, and then hands off to shot-specifier for per-shot direction, model routing, and prompt manifests. Use when analysing stories, scripts, or prose to create production-ready scene breakdowns with full visual asset pipelines. Also trigger when the user mentions "scene breakdown", "shot list", "character bible", "location bible", "continuity inventory", "reference images", "storyboarding", or any request to prepare narrative material for AI video generation. This skill expects access to an image-generation MCP and vision capabilities.

Scene Inventory Extractor

Systematic workflow for extracting comprehensive scene inventories from narrative source material and preparing every visual asset required for AI video generation. Continuity is a first-class deliverable: extract story-state, dressing-state, and object-state before writing prompts so a crew can reset scenes accurately when production shoots out of order.

Execution Context

This skill is designed for a command-line agent with:

• nanobanana image generation MCP (generate_image, edit_image, character_consistency, and multi_image_fusion as needed), with every image call explicitly using model: gemini-3-pro-image-preview
• Vision capabilities (ability to inspect generated images for consistency)
• File system access (structured output directories)

All image generation and verification runs silently — no user confirmation gates during generation phases. Halt only on consistency failures that require human judgement. If gemini-3-pro-image-preview is unavailable through the nanobanana MCP, or if it cannot accept the reference images or character-consistency images required by the current phase, STOP and report the blocker. Do not switch to another image model.

Workflow Overview

Skill chain: scene-inventory-extractor-v2 prepares the scene pack and image assets. Hand off to shot-specifier when per-shot direction, storyboard generation, and prompt manifests are needed. Hand off to video-generator only after prompt files, frame paths, model routing, and generation strategy fields exist and the goal is to submit jobs through the Higgsfield MCP.

Read order for reference files: Before starting Phase 2, read references/cinematography-specification.md. Before Phase 8, read references/continuity-inventory.md. Before Phase 11, read references/reference-image-guide.md. The consistency verification procedure in Phase 13 is defined in references/consistency-verification.md. The prompt keyword library format is defined in references/prompt-keyword-library.md. Downstream skill: This skill stops at Phase 13. When full per-shot storyboarding, detailed actor/camera/lighting direction, model routing, prompt flattening, audio generation preferences, and asset-pipeline management are required, hand off to the shot-specifier skill. That skill takes the scene inventory and reference images produced here as its input. When video generation itself is required, use video-generator; this skill does not call Higgsfield directly.

Phase 1: Source Analysis

Read the source material twice. First pass: absorb narrative arc, tone, world. Second pass: extract with pen in hand.

First-Pass Questions

• What is the dominant visual mood? Name it.
• Where does the camera naturally want to be? (Observational? Intimate? Surveillance?)
• What objects carry narrative weight?
• What spaces define the world?
• Who moves through this world and how do they move?
• What is the light doing? (Time of day, weather, artificial sources)
• What is the sound doing? (Ambient beds, punctuations, silence)

Phase 2: Creative Pillars

Prerequisite: Read references/cinematography-specification.md before this phase.

Extract and name the visual, storytelling, and cinematographic approaches.

2.1 Visual Aesthetic

**Name:** {Evocative 2–4 word title}
**Definition:** {One sentence capturing look and feel}

Specify: Palette (3–5 named colours), Lighting rules, Texture rules, Camera grammar, Warmth rules. See templates/scene-inventory-template.md for field structure.

2.2 Storytelling Style

**Name:** {Evocative 2–4 word title}
**Definition:** {One sentence capturing narrative approach}

Specify: Narrative rules, Rhythm rules, Scale rules.

2.3 Cinematography Specification

This is new territory relative to pure script breakdown. Define the physical and post-production characteristics of the image pipeline:

Full specification format and examples are in references/cinematography-specification.md. Use this resolution contract consistently across extractor, shot-specifier, and video-generator: Target resolution is always explicit pixel dimensions, while resolution-parameter carries any provider preset label.

2.4 Prompt Keyword Library

Prerequisite: Read references/prompt-keyword-library.md before this sub-phase.

Immediately after completing the cinematography specification, derive a project-level prompt keyword library: a canonical vocabulary of adjective phrases and art-direction terms that reliably translate each style parameter into language video and image generation models respond to consistently.

This library is infrastructure, not decoration. Every model-routed video prompt later written by shot-specifier must draw from it rather than re-inventing style language shot by shot. Inconsistent style vocabulary across shots produces visible tonal drift even when the underlying spec is correct.

Write the library as a section in the scene inventory document and also as a standalone file at {project_name}_prompt_keywords.md.

### Prompt Keyword Library

* **Global style phrase:** {1–2 sentences encoding filmstock, grain, colour process}
  * Example: "35 mm print, Kodak Vision3 250D characteristics, organic fine-to-medium grain,
    accurate daylight balance, slightly desaturated shadows, gentle highlight rolloff"

* **Per-location-type vocabulary:**
  * {Location type}: {canonical phrase set for colour, light, texture, atmosphere}
    * Example exterior: "flat pewter Atlantic overcast, wet peat surfaces, no hard shadows,
      lifted blacks, fence posts leaning in wind"
    * Example interior warm: "warm amber grow-light, artificial June, high humidity haze,
      strawberry scarlet at full vibrancy"

* **Per-lighting-condition vocabulary:**
  * {Condition name}: {phrase set}
    * Example pre-dawn: "sodium-orange practicals, deep grey-blue surround, rain halos
      around light sources, near-dark ground plane"

* **Global negative constraints:** {what must never appear in prompts}
  * Example: "no trees, no dual carriageways, no US road markings, no hard shadows,
    no blue sky, no bright saturated colours except strawberry scarlet"

* **Selective saturation rules:** {objects exempt from global desaturation}
  * Example: "strawberry scarlet is exempt from the global 15–20% desaturation pass;
    push selectively against every other desaturated element"

* **POV-specific overrides:** {e.g. machine-vision shots that require no grain}
  * Example drone POV: "digital-flat, no grain, deep focus throughout, gimbal-stabilised,
    machine vision — no organic camera movement"

Phase 3: Narrative Spine

Document timeframe, structural approach, operational themes (4–6), and key turnpoints (3–7). Themes must be operational: "Signature authority: who gets to update 'reality'" not "The loneliness of modern existence."

Phase 4: Character Bible

For each character, extract using the appropriate template from templates/character-template.md (primary / functional / interface / voice-only / collective).

Addition — Reference Image Specification per character:

For every character who appears on screen, append:

* **Reference image requirements:**
  * Primary: {Full body, neutral pose, front-facing, white background}
  * Face detail: {Head-and-shoulders, ¾ angle}
  * Expression set: {List key expressions needed by the story}
  * Wardrobe variants: {Each distinct outfit as a separate ref}
  * Action poses: {Any signature physical action}

Match detail to narrative weight: primary characters get full sets; functional characters get primary + face; collectives get a representative figure.

Also record a continuity state chain per recurring on-screen character: wardrobe by scene, carried items by scene, body-state continuity, and pocket / hand continuity risks. "Same outfit" is not enough when carried objects or physical condition change.

Phase 5: Locations Bible

Critical: Locations receive the most comprehensive reference-image coverage. Each location that appears in more than one scene, or whose atmosphere shifts across the story, requires multi-angle, multi-condition scouting.

For each location, extract using templates/location-template.md.

Scouting Specification

For each location, define:

The scouting matrix is: angles × (lighting conditions × weather conditions × seasonal/narrative states). Generate the full cross-product only where the story actually visits that combination. Do not generate speculative combinations.

For recurring locations, add a continuity chain across appearances: first-appearance state, subsequent appearance changes, and non-negotiable anchors that must never drift.

Phase 6: Props Bible

For each significant prop (passes the significance test: narrative weight, multi-scene, character-defining, symbolic, or physically interacted with on camera), extract physical description, narrative function, sequence appearances.

For continuity, also record custody, state progression, set-down / pickup moments, and scene exit status.

6.1 Recurring Visual Elements

Also identify Recurring Visual Elements: any object, fixture, interface, machinery, set dressing element, furniture layout, or repeated environment component that appears in more than two shots and has a specific enough appearance that the audience would notice if it changed. These elements do not need full Props Bible custody/state entries unless they are handled or narratively transformed, but they do need locked reference images.

Treat recurring visual elements as reference-required even when they feel like location dressing. Examples include monitor-bank layouts and screen colours, inspection robots, grow-light strip configurations, cargo pod designs, storage cabinet designs, recurring signage clusters, and a named character's workstation or console layout.

For each recurring visual element, record:

* **Recurring visual element:** {Name}
  * Location / set: {Where it belongs}
  * Appearance lock: {Shape, layout, colour, screen state, arrangement, scale}
  * Appears in shots/scenes: {S01_SH001, S01_SH003, ...}
  * Reference image requirement: locked in Phase 11
  * Reference file: refs/recurring-elements/{name}/primary.png
  * Must pass as referenceImagePaths in: {All shots where visible}

Addition — Reference Image Specification per prop:

* **Reference image requirements:**
  * Reference priority: required-before-Phase-12 / incidental
    (required-before-Phase-12: prop appears prominently on screen across multiple shots
     and its visual identity is story-critical — a named vehicle, weapon, device, or
     object whose appearance the audience will track. Generate before any location or
     scene image that includes it. incidental: prop appears briefly or in background;
     standard Phase 11 timing applies.)
  * Primary: {Object on white background, ¾ angle}
  * Detail: {Any specific detail the camera will see in ECU/INS}
  * In-context: {Object in its typical environment, if context matters}
  * State variants: {If prop changes condition across the story}

Phase 7: Scene Inventory

For each scene (new scene on location change, significant time passage, POV shift, or mode change):

#### SC-{XX}

* **Scene ID:** SC-{XX}
* **Location:** {Location name}
* **Time:** {Time of day; weather; season}
* **Lighting condition key:** {Maps to scouting matrix entry}
* **Characters present:** {Who}
* **Objective / tension:** {What's at stake}
* **What changes:** {State shift by scene end}
* **Key sensory notes:** {Smell/sound/temperature}
* **Transitions in/out:** In: {from}. Out: {to}.
* **Continuity dressing notes:**
  * Fixed location anchors: {architecture, furniture, installed fixtures}
  * Recurring visual elements: {monitor banks, robots, light strips, cabinets, console layouts, signage clusters}
  * Movable dressing: {objects that can shift position}
  * Character-carried items: {by character}
  * Consumables / depletion states: {food, drink, cigarettes, fuel, paper stacks}
  * Weather / dirt / damage state: {mud, blood, sweat, rain, soot, wrinkles}
  * Reset-sensitive details: {what must match across coverage and return visits}

Phase 8: Continuity Inventory

Prerequisite: Read references/continuity-inventory.md before this phase.

Write a separate continuity deliverable at {project_name}_continuity_inventory.md. Continuity extraction is pre-generation story/state tracking, not post-generation image QA.

Assume scenes will be shot out of order. Extract continuity so that a separate crew, working days later, can restore the exact character, prop, and dressing state without re-reading the source material.

Phase 9: Shot Lists

Create shot tables for each sequence. Use the expanded format from templates/shot-list-template.md, which now includes cinematography columns.

Duration Budget (do this first)

Before decomposing each scene into individual shots, establish a duration budget:

1. Estimate the total desired screen time for the scene in seconds.
2. Divide into clips: clips must be 4, 6, or 8 seconds each.
3. Record the budget before writing individual shot rows — it prevents collapsing multi-beat scenes into a single over-stuffed clip.

* **Scene duration budget:** {total seconds}
* **Clip count:** {N clips}
* **Per-clip targets:** {e.g. SH001: 8s, SH002: 6s, SH003: 4s}

Common error: SC-11-style launch sequences feel like one shot but contain at least three distinct clips (pre-launch static, rotor spin-up and lift, forward flight receding). Budget before decomposing, not after.

Expanded Shot Table Columns

Per-Shot Frame Specification

For each shot, also record (these feed Phase 12):

* **Start frame:** {Framing + visible content + subject state}
* **End frame:** {Framing + visible content + subject state}
* **Key frames (if any):** {Intermediate states the interpolation must pass through}
* **Interpolatable change:** {What changes between start and end — position, pose, state, composition}
* **End-frame derivation:** edit-from-start / generate-new (see decision table in references/reference-image-guide.md)

[CRITICAL] The end frame must show interpolatable change from the start frame: subject position/pose, subject state, or composition shift. Subtle-only changes (lighting, background) while the subject stays static cause unnatural video motion.

Phase 10: Thematic Image Plan

Identify 8–12 key images capturing narrative beats. Specification format unchanged — see template. These images serve double duty: editorial illustration and generation quality-gate (if the pipeline cannot produce a convincing thematic image, it flags a capability gap before committing to full shot generation).

Phase 11: Reference Image Generation

Prerequisite: Read references/reference-image-guide.md before this phase.

Generate all reference images silently. Order matters:

11.1 Style Anchor

Generate 1–2 style-anchor images that establish the global look (filmstock, grain, palette, lighting). These become the visual-style reference for everything that follows.

11.2 Character References

For each character, generate in order:

1. Primary reference (no prior refs; full visual-style spec in prompt + white bg)
2. Additional angles/expressions/wardrobe (primary ref as input reference)

11.3 Prop References

Order rationale: Props are generated before locations because location shots often contain key props in frame. Without a locked prop reference, each location image will independently invent the prop's appearance — producing the failure mode where the same aircraft looks like three different vehicles across a sequence.

For each prop with reference priority: required-before-Phase-12:

1. Primary reference (white background, ¾ angle, style spec in prompt)
2. Detail / state variants (primary ref as input)

Then, after locations are complete (11.5), generate refs for incidental props in the same pattern.

11.4 Recurring Visual Element References

Generate a locked primary reference for every recurring visual element identified in Phase 6.1. Use the element's normal in-context environment unless a white-background plate would better define the geometry.

Generate these before location scouting references whenever the element appears inside a location reference. Otherwise the location reference can bake in one invented version of the element while later shot frames invent another.

11.5 Location Scouting References

For each location, generate across the scouting matrix:

1. Primary establishing shot (no prior refs; full style spec in prompt)
2. Additional angles (primary ref as input)
3. Lighting/weather/condition variants (primary ref as input; describe the changed conditions)

Generation Rules

• Use the nanobanana MCP for every reference-image call and pass model: gemini-3-pro-image-preview explicitly.
• If gemini-3-pro-image-preview cannot run the requested operation with the required primary references, character references, or multi-image inputs, STOP the workflow and report the unavailable capability. Do not continue with a fallback model.
• Every prompt ends with "no text, no watermarks, no logos, no labels, no annotations"
• Additional refs always use the primary ref as a reference image input
• Prompts for additional angles carry forward the style spec in abbreviated form
• All images use the aspect ratio defined in the cinematography specification
• File naming: ref_{category}_{name}_{variant}.png
  • e.g. ref_char_miette_primary.png, ref_loc_control-room_dusk-rain.png

11.6 Video Role Manifest

After generating all reference images, produce a video role manifest that declares the intended role of each image when used as a video generation input. This is distinct from the reference image manifest (which tracks what exists); the video role manifest tells downstream tools and the shot-specifier skill how to use each image.

## Video Role Manifest

| Ref ID | File | Video Role | Used In Shots | Notes |
|--------|------|------------|---------------|-------|
| LOC-launch-01 | refs/locations/launch-strip/low-pre-dawn-rain.png | start_image | S11_SH001 | Pre-launch static |
| LOC-launch-03 | refs/locations/launch-strip/gannet-vertical-lift.png | end_image | S11_SH002 | Mid-lift anchor |
| PROP-gannet-01 | refs/props/gannet-uav/primary.png | image | S11_SH001, S11_SH002 | Subject consistency ref |
| RVE-control-monitors-01 | refs/recurring-elements/control-room-monitor-layout/primary.png | image | S03_SH001, S03_SH004 | Monitor layout and screen colour lock |
| CHAR-switch-01 | refs/characters/switch/primary.png | image | S08_SH001 | Identity anchor |
| style/style_anchor_01.png | refs/style/style_anchor_01.png | image | all | Global style ref |

Video role values:

• start_image — anchors the first frame of the clip
• end_image — anchors the last frame of the clip
• image — subject/style consistency reference (visible in clip but not a frame anchor)
• video — reference video for motion style
• audio — reference audio track

Phase 12: Shot-Frame Generation

Prerequisite: Phase 11 complete. All reference images available.

Pre-Generation Reference Check (per shot)

Before generating any frame for a shot, answer these questions explicitly:

1. Does a canonical reference image exist for every named character present in this shot?
2. Does a canonical reference image exist for every required-before-Phase-12 prop visible in this shot?
3. Does a locked reference image exist for every recurring visual element visible in this shot?
4. Does a canonical reference image exist for the specific location variant (angle × lighting condition) this shot requires?

If any answer is no, generate that reference now using the Phase 11 procedure before proceeding. Do not skip this check. The failure mode it prevents: a scene frame generated before the prop, recurring visual element, or location reference exists, where the model independently invents monitor layouts, fixture arrangements, robots, cabinets, or prop appearance — producing visibly different objects or set dressing across shots.

For every shot in every sequence, generate:

1. Start frame — using appropriate character, location, and prop references
2. End frame — derived from start frame (edit or generate, per shot spec)
3. Key frames — any intermediate states specified in the shot list

Start Frame Generation

• Character-centric shots: Tool: nanobanana MCP character_consistency
• Environment or prop-led shots: Tool: nanobanana MCP generate_image
• Model: gemini-3-pro-image-preview
• referenceImagePaths for character-centric shots: character identity reference first, then the location ref matching lighting/weather, required prop refs, and the style anchor when available; include any visible recurring visual element refs
• referenceImagePaths for environment or prop-led shots: location ref matching lighting/weather, required prop refs, recurring visual element refs, and the style anchor when available; include character refs only if a named character is visible and identity must be constrained
• Prompt includes: visual style (brief), scene environment, framing, visible content, subject appearance + outfit
• Draw style vocabulary from the prompt keyword library produced in Phase 2.4
• Aspect ratio: from cinematography specification
• Prompt ends with: "no text, no watermarks, no logos, no annotations"

End Frame Generation

If edit-from-start:

• Tool: nanobanana MCP edit_image
• Model: gemini-3-pro-image-preview
• referenceImagePaths: [start_frame, only the Phase 11 refs needed to describe the intended delta]
• Prompt: "Edit this image: {changes only}" — do NOT repeat unchanged elements
• Use this path whenever the end frame derives from the start frame; the start frame carries character, location, prop, and style consistency forward naturally.

If generate-new:

• Tool: nanobanana MCP generate_image
• Model: gemini-3-pro-image-preview
• References: [start_frame (as scene ref), relevant Phase 11 refs]
• Prompt includes: visual style, end-frame framing + visible content, subject end state, "Same location/environment as reference"

Key Frame Generation

For each specified key frame, choose the tool by the frame's dominant continuity risk:

• Character-centric key frame: use nanobanana MCP character_consistency with model: gemini-3-pro-image-preview. Put the character identity reference first in referenceImagePaths, followed by start_frame, matching location ref, required prop refs, and style anchor when available.
• Environment or prop-led key frame: use nanobanana MCP generate_image with model: gemini-3-pro-image-preview. Set referenceImagePaths to [start_frame, matching location ref, required prop refs, recurring visual element refs, style anchor when available].
• Key frame derived from the start frame by pose, expression, object state, or minor camera adjustment: use nanobanana MCP edit_image with model: gemini-3-pro-image-preview. Set referenceImagePaths to [start_frame, needed Phase 11 refs] and describe only the change.

Do not use generate_image for a character-centric key frame just by adding the character reference to a mixed reference pool. Use character_consistency so identity anchoring is the operation's primary constraint.

File Naming

shot_{shot_id}_{frame_type}.png — e.g. shot_S01_SH003_start.png, shot_S01_SH003_end.png, shot_S01_SH003_key01.png

Phase 13: Consistency Verification

Prerequisite: Read references/consistency-verification.md.

After generating all shot frames, run a vision-based consistency pass. This pass is not informational. Its findings are work items for the agent: resolve BLOCK issues by regenerating or correcting the offending frame before handoff, and either resolve WARN issues or carry them forward with explicit shot-specifier instructions.

For each shot:

Output a Consistency Report listing every flagged issue with severity (BLOCK / WARN), the shot ID, the check that failed, and a description. BLOCK-level issues must be resolved by regenerating the offending frame before handoff. WARN-level issues are not passive notes: resolve them when the fix is clear; otherwise write the concrete constraint shot-specifier must inject into the affected shot.

Handoff Package

After Phase 13 passes with no unresolved BLOCK issues, compile the final scene pack and handoff notes for shot-specifier. Do not assemble video prompts in this skill. Do not choose final video models. Do not flatten prompts. Do not call Higgsfield.

Compile the final scene inventory document using templates/scene-inventory-template.md.

Document Structure

1. Header (title, version, date, logline, scope)
2. Creative Pillars (including Cinematography Specification and Prompt Keyword Library)
3. Narrative Spine
4. Character Bible (with reference image manifest)
5. Locations Bible (with scouting matrix and reference image manifest)
6. Props Bible (with reference image manifest)
7. Scene Inventory
8. Continuity Inventory
9. Style Frames Audit
10. Storyboard Specification (sequence map + shot lists with frame specs and duration budgets)
11. Thematic Image Plan
12. Consistency Report
13. Shot-frame asset manifest
14. Handoff notes for shot-specifier
15. Continuity & Anachronism Log
16. Asset Manifest (all generated images, organized by category)

Output Files

The extractor-to-shot-specifier contract has one formal handoff artifact: the Phase 13 handoff package. Downstream consumers should read these output keys from that package rather than expecting an additional extractor handoff step.

{project_name}/
├── {project_name}_scene_inventory.md
├── {project_name}_continuity_inventory.md
├── {project_name}_prompt_keywords.md
├── {project_name}_recurring_visual_elements.md
├── refs/
│   ├── style/
│   ├── characters/
│   ├── locations/
│   ├── recurring-elements/
│   └── props/
├── shots/
│   └── {shot_id}/
│       ├── start.png
│       ├── end.png
│       └── key{NN}.png (if any)
└── reports/
    ├── phase_13_handoff.md
    └── consistency_report.md

Reference Files

Templates

Best Practices

Read Twice, Extract Once. Complete a full read before extraction.

Operational Language. Concrete, filmable. "Jaw tightens" not "feels tension."

Constraint-First. For every element: what must NOT appear, as well as what must.

Continuity First. Extract narrative and physical continuity before prompt writing. Post-generation consistency verification is a later QA step, not a substitute.

Out-of-Order Production Reality. The continuity inventory must let another crew restore exact dressing, object state, and carry chains days later without re-reading the source.

Mundane Objects Count. Track all handled objects, not just hero props. If a mug, pen, phone, keyring, utensil, paper stack, cigarette, badge, bag, or desk object can move or change state, it belongs in continuity tracking.

Budget Before Decomposing. Establish a clip-duration budget for each scene before writing individual shot rows. Multi-beat scenes collapsed into a single oversized clip produce rushed, incoherent video. Three clips of 6s each is almost always better than one clip of 18s.

Prompt Vocabulary is Production Infrastructure. The prompt keyword library is not optional polish — it is the mechanism that keeps visual style consistent across dozens of independently generated clips. Style vocabulary invented ad hoc per shot produces tonal drift. Build the library once in Phase 2.4 and pass it to shot-specifier.

Pass Continuity Constraints Downstream. Location bibles contain negative space rules, colour exclusions, and signature prop requirements. These must be included in the handoff notes so shot-specifier can carry them into each video prompt; reference images enforce visual consistency but cannot enforce textual rules ("no trees", "left-hand traffic", "no blue sky outside").

Evocative Naming. "Bruise Blue" not "#2B4F6E". "Kodachrome Sunday" not "warm LUT."

Function Over Description. State narrative function for every element.

Sound as Character. Audio beds, punctuations, and silence matter as much as visuals.

Reference Images are Non-Negotiable. Never generate shot frames without reference images. Never generate additional refs without using the primary as a reference input.

References First, Always. Lock canonical reference images for every named character, every required-before-Phase-12 prop, and every key location before generating any composite scene or storyboard frame. The generation order is: style anchor → characters → required-before-Phase-12 props → locations → incidental props. Never generate a location image that contains a named prop until that prop's primary reference is locked. Skipping this produces the category error where the same object — a vehicle, a device, a named weapon — looks like a completely different thing in every shot it appears.

Start and End Frames for Every Shot. No exceptions. The video model needs both anchors to interpolate convincingly.

Verify Before Handoff. Run the consistency pass (Phase 13) before handing off to shot-specifier. The agent must action the consistency report before handoff: fix BLOCK findings, resolve fixable WARN findings, and turn any remaining WARN findings into explicit downstream constraints. Catching a face swap now costs one regeneration; catching it after video generation costs the entire shot.