refactor: extraer el router HTTP de LLM

[IAnMove]

Qué

Extrae el dominio HTTP de /api/v1/llm fuera de _launch_runtime.py hacia app/routers/llm.py, imitando el montaje de Comics (create_comics_router + segundo include_router para no romper ordinales).

Contrato (11 rutas)

create_llm_router (ordinales 80–87, misma posición que el primer /api/v1/llm):

• GET /api/v1/llm/status
• POST /api/v1/llm/load
• POST /api/v1/llm/unload
• GET /api/v1/llm/models
• GET /api/v1/llm/stream-status
• POST /api/v1/llm/generate
• POST /api/v1/llm/test
• POST /api/v1/llm/write-song

create_llm_prompt_router (ordinales 89–91, segundo include_router justo después de Director generate-music):

• POST /api/v1/llm/plan-h3-windows
• POST /api/v1/llm/enhance-prompt
• POST /api/v1/llm/describe-image

route_table.json no cambia ordinales, method, path, endpoint ni status. Solo source pasa a app/routers/llm.py en esas 11 rutas.

Cómo se conservan los ordinales

scripts/architecture_contracts.py expande cada factory en su api.include_router(...). El primer montaje sustituye el bloque control/generate/song-writer; el segundo se coloca justo donde estaban plan-h3 / enhance / describe, después de POST /api/v1/director/generate-music (ordinal 88, sigue en launch).

Qué permanece en launch

• /api/v1/director/* (incluido director_generate_music, plan-prompts, pipeline). Director es Paso 5 y está fuera de alcance.
• _ensure_llm_loaded (lo usan CivitAI, Director, Story Lab, Comics, describe-refs, …).
• _enhance_with_wangp (WanGP/offload; se inyecta en el router).
• _comic_writing_llm, _effective_llm_routing, _llm_provider_credentials.

Director generate-music importa _parse_song_output y los fallbacks ACE-Step desde routers.llm. llm_service.py no se parte (plan 6.b no es este ticket).

Inyección (sin WanGP en el router)

• Control/generate/song: get_services_config, effective_llm_routing, llm_provider_credentials, llm_default_device, default_llm_repo, ensure_llm_loaded, comic_writing_llm.
• Prompt tools: lo anterior + public_llm_providers, get_model_def, get_lora_dir, get_cached_hardware, get_enhancer_enabled, enhance_with_wangp.

El router no importa wgp ni _launch_runtime.

Tests

python -m compileall -q app/routers/llm.py app/services/llm_service.py
python scripts/architecture_contracts.py
python -m pytest -q tests/test_architecture_contracts.py tests/test_llm_router.py tests/test_minimax_song_writer_prompt.py tests/test_h3_window_planner.py tests/test_minimax_h3.py::TestMiniMaxH3Definition::test_h3_enhance_path_preserves_context_ir_contract

No hace falta GPU ni create_app().

Los tests que rebanaban launch para song-writer / plan-h3 / enhance se retargetean a app/routers/llm.py. test_minimax_song_writer_prompt.py deja de ser lector de launch (sale del wire inventory). tests/test_llm_router.py cubre superficie path/method/endpoint y validación 400.

Fuera de alcance

• No se toca /api/v1/director/*, stories, series, audio, video-editor.
• No se toca ui/src/stores/useStore.ts, WanGP, app/wgp.py, app/models/**, launchers, Pinokio.
• No se toca docs/development/SLICE_QUEUE.md (nota post-merge: marcar el router LLM como landed en «Backend by domain»).
• Un solo commit revertible. Sin features nuevas.

No mezclar hasta CI verde.

---

Note

^{Cursor Bugbot is generating a summary for commit 91e80ef. Configure here.}

[github-actions]

PR Review — Loreframe Studio

Risk: medium
Scope: 8 file(s); +928/-651

Automated review from scripts/analyze_pr.py. This is a heuristic pass (no LLM) so humans still own the merge decision.

Findings

• medium — Large pull request
  928 additions / 651 deletions. Reviewers will have an easier time with smaller, focused PRs.
• medium — Very large file change (app/routers/llm.py)
  app/routers/llm.py adds 759 lines. Consider splitting the PR.

Changed files

• added: app/routers/llm.py, tests/test_llm_router.py
• modified: app/_launch_runtime.py, tests/fixtures/architecture_wire_inventory.json, tests/fixtures/route_table.json, tests/test_h3_window_planner.py, tests/test_minimax_h3.py, tests/test_minimax_song_writer_prompt.py

CONTRIBUTING checklist

• python scripts/verify_clean_repo.py
• python -m compileall -q app/services app/launch.py scripts
• cd ui && npm run build if the UI changed
• No weights, CivitAI sidecars, or generated guides
• Stays local-first (no required accounts / telemetry)

Posted by the repo PR review workflow. Re-runs on each push to the PR.

[github-actions]

Code health

Metric	Value
Production LOC	235,061
Production files	509
Test LOC	64,413
Functions measured	14,750
Functions complexity ≥ 15	773
Maximum complexity	667

Markdown, JSON catalogs and tests are out of this table. Only app/ runtime + ui/src TS/JS count.

Most complex functions

Complexity	Where
667	app/wgp.py:7157 generate_video
374	ui/src/stores/useStore.ts:4426 Async method 'startGeneration'
356	app/_launch_runtime.py:23154 _run_generation
308	app/wgp.py:12274 generate_video_tab
272	ui/src/components/Sidebar/SceneAnimatorPanel.tsx:469 Function 'SceneAnimatorPanel'
266	ui/src/stores/useStore.ts:9164 Async method 'loadSettingsFromOutput'
258	app/services/director/planners/short_film.py:3433 ShortFilmPlanner._plan_story_driven
254	app/services/director_pipeline.py:13539 _run_video_generation
245	app/services/director_pipeline.py:7663 _run_pipeline
241	ui/src/features/agent/agentActions.ts:1077 Function 'parseAction'
227	ui/src/features/agent/agentActions.ts:2672 Async function 'executeAgentActions'
226	app/services/director_pipeline.py:6492 update_comic_preview

Trend vs baseline

Metric	Δ
Production LOC	+1,396
Test LOC	+599
Functions ≥ 15	-1
Maximum complexity	+0

Warnings

• production LOC increased by +1,396
• complexity hotspot ui/src/stores/useStore.ts rose 373 -> 374
• complexity hotspot app/services/model3d_service.py rose 51 -> 54
• complexity hotspot ui/src/components/MainContent/TabFilter.tsx rose 19 -> 23
• complexity hotspot ui/src/features/agent/AgentAssistantPanel.tsx rose 19 -> 23
• hotspot ui/src/features/agent/agentActions.ts increased by +3 lines
• hotspot ui/src/features/video-editor/VideoEditorPanel.tsx increased by +10 lines
• hotspot ui/src/types/index.ts increased by +3 lines
• hotspot app/services/model3d_service.py increased by +58 lines
• hotspot ui/src/features/agent/capabilityRegistry.ts increased by +2 lines
• hotspot ui/src/features/agent/applicationAdapters.ts increased by +7 lines

Ratchet passed.