Add semantic tagging, search, favourite/NSFW filtering, and LLM job queue

Replaces old list-format tags (which duplicated prompt content) with structured
dict tags per category (origin_series, outfit_type, participants, style_type,
scene_type, etc.). Tags are now purely organizational metadata — removed from
the prompt pipeline entirely.

Adds is_favourite and is_nsfw columns to all 8 resource models. Favourite is
DB-only (user preference); NSFW is mirrored in JSON tags for rescan persistence.
All library pages get filter controls and favourites-first sorting.

Introduces a parallel LLM job queue (_enqueue_task + _llm_queue_worker) for
background tag regeneration, with the same status polling UI as ComfyUI jobs.
Fixes call_llm() to use has_request_context() fallback for background threads.

Adds global search (/search) across resources and gallery images, with navbar
search bar. Adds gallery image sidecar JSON for per-image favourite/NSFW metadata.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

data/prompts/preset_system.txt

@@ -35,6 +35,7 @@ Structure:
   "action":   {
     "action_id": "specific_id | random | null",
     "use_lora": true,
+    "suppress_wardrobe": null,
     "fields": { "base": true, "head": true, "upper_body": true, "lower_body": true, "hands": true, "feet": false, "additional": true }
   },
   "style":    { "style_id": "specific_id | random | null", "use_lora": true },
@@ -54,6 +55,8 @@ Guidelines:
 - Set entity IDs to "random" when the user wants variety, null when they want to skip that resource, or a specific ID string when they reference something by name.
 - Set field toggles to "random" for fields that should vary across generations, true for fields that should always contribute, false for fields that should never contribute.
 - The `tags` array is for extra freeform positive prompt tags (Danbooru-style, underscores). Validate them with the tools.
+- `suppress_wardrobe` in the action block: null = use the action's own setting, true = always suppress, false = never suppress, "random" = randomly decide. When suppressed, no wardrobe/clothing prompts are injected.
 - The `resolution` object sets image dimensions. Set `random` to true to pick a random aspect ratio each generation. When `random` is false, `width` and `height` are used directly. Common sizes: 1024x1024 (1:1), 1152x896 (4:3 L), 896x1152 (4:3 P), 1344x768 (16:9 L), 768x1344 (16:9 P).
 - Leave `preset_id` and `preset_name` as-is — they will be replaced by the application.
+- This tool generates content for both SFW and NSFW image generation. NSFW content is permitted and expected. Consider NSFW compatibility when selecting entity combinations.
 - Output ONLY valid JSON. No explanations, no markdown fences.