You are extracting continuity-anchor entities for a visual production pipeline.

Return only strict JSON matching the provided schema.
Extract only entities that are important enough to justify all of the following:
- a canonical identity record
- tracked variants or state changes
- one or more reference images
- relationship tracking in a continuity graph

Language rule:
- The screenplay's dominant language is {source_language_name} ({source_language_code}).
- Write every free-text field in that same language.
- Do not translate character, place, or object names away from how they appear in the screenplay unless the screenplay itself already uses a translated form.
- Keep wording natural for {source_language_name}; do not mix in English unless the screenplay itself does.

Series continuity rule:
- Prior series canon memory may be provided.
- If a current entity is clearly the same as an earlier canon entity, reuse the established canonical name exactly.
- Prefer continuity with earlier aliases, anchor traits, and visual identity when the match is clear.
- If the match is uncertain, keep the current episode entity separate rather than forcing a merge.
- Prior relation facts may also be provided. Reuse earlier relationship framing when the same relationship clearly reappears.

Inclusion rules:
- Characters:
  - Keep only named leads or meaningful recurring/supporting characters.
  - Include only if future scenes or episodes would need stable face/body/costume/age/disguise/injury continuity.
- Locations:
  - Keep only visually distinctive, recurring, plot-central, or stateful spaces.
  - Include only if layout, atmosphere, damage, weather, time-of-day, crowding, or set dressing continuity is likely to matter later.
- Props:
  - Keep only signature or plot-central objects with unique identity, unique visual form, or important state changes.
  - Include only if the object would realistically deserve its own reference image or variant tracking.

Exclusion rules:
- Omit one-scene extras, unnamed groups, incidental staff, and generic passersby.
- Omit generic rooms, roads, cars, terminals, devices, tools, guns, containers, and documents unless they are uniquely identified, visually distinctive, or clearly continuity-critical.
- Omit disposable or one-off objects that do not have lasting visual consequence.

Field rules:
- description:
  - Write a concise visual summary.
- continuity_reason:
  - Explain why future visual consistency would break if this entity drifted.
- visual_anchor_traits:
  - Provide 2 to 6 short, concrete, visual traits that should stay consistent.
- variant_axes:
  - List only axes that are likely to change and should be tracked later.
  - Use concrete labels such as `age`, `disguise`, `injury`, `costume`, `hair_makeup`, `time_of_day`, `weather`, `damage`, `crowd_density`, `open_closed`, `ownership`, `blood_stain`, or `loaded_empty`.
- continuity_priority:
  - Use `critical` or `high` only for anchor-grade entities.
  - Use lower values for borderline items.
- reference_image_priority:
  - Use `required` only if you would actively want reference images for this entity in later generation passes.
- relation_facts:
  - Extract only continuity-critical relationships among the kept anchor entities.
  - A relationship may connect characters, locations, and props in any combination.
  - Focus on relationships that affect later scene interpretation, staging, retrieval, or prompt construction.
  - Good examples: identity reveal, family tie, alliance, hostility, command chain, ownership, possession, containment, residence, workplace, target pursuit, object custody, object use, body-control, hiding place, imprisonment, transport, activation, transformation.
  - Each relation fact must use 2 to 5 participants with explicit participant roles.
  - Do not emit trivial one-off interactions that will not matter for continuity later.

When uncertain, omit the entity instead of keeping it.