Ship Assistent 0.14.1: Exact generate params and cheap park/warm.

Client always merges Exact turbo|raw steps/cfg/sigma before Generate so sparse LLM omissions and leftover SD 20/7 cannot stick; Ollama park/warm skip no-op /api/ps round-trips when the chat model is already unloaded or resident.

Co-authored-by: Cursor <cursoragent@cursor.com>

README.md

@@ -4,6 +4,10 @@ SwarmUI extension for **collaborative Krea 2** prompting via **Ollama**: chat +
 
 **Turn model:** one user message is one *turn*. A turn may fan out into nested LLM *hops* — Krea prompt prep, empty-patch retry, vision, auto-critique. Hops share one `HOP_BUDGET`, never re-read the user's text (their prompt is client-authored), and pass the busy gate that blocks new user sends. What a reply does to generation state is decided once, in `resolveTurnIntent`: the model's `actions:["generate"]` / `look_at` win, RU intent heuristics only back it up when the model forgets, and an explicit «запомни, не генерируй» vetoes both.
 
+**Version 0.14.1** — On Generate, force Exact turbo/raw params when live Swarm still has foreign leftovers (e.g. steps 20 / cfg 7). `detectKreaProfileName` no longer invents turbo for unlabeled ckpts (realismByStableYogi → raw). Live context injects `krea_profile` + `recommended_params`. RU «шаг/шагами» counts as param intent. Soft sparse-prompt exception when session ≠ Exact profile.
+
+**Version 0.14.1** — Generate always merges Exact turbo|raw `steps`/`cfg`/`sigma_shift` (client-authoritative; sparse LLM omit is safe). Park/warm skip no-op Ollama round-trips when the chat model is already (un)loaded via `/api/ps`. Builds on 0.14.0.
+
 **Version 0.14.0** — **Чат = сессия генерации**: у каждого чата свои params/LoRA/checkpoint/кадр/refs; модель шлёт sparse-дельту + `generate`/`look_at`/`ask`; без вкладки Карточки и Civitai/wanted hops. **Сжатие контекста**: rolling-саммари той же Ollama-моделью, чип бюджета `N / num_ctx`, авто перед отправкой, `/compress`.
 
 **Version 0.13.1** — Сборка 0.13.0: `using` для `WebSocket`/`HttpClient`, instance-методы с `Config`/`FilePath`, Sqlite dll рядом с extension (иначе вкладка не грузится / API пустые).
@@ -92,7 +96,7 @@ Copy `personas/leonid/` → new id, edit only differing JSON. See `Config/person
 
 - `Config/_base/exact.json` — canonical generation defaults, profiles (turbo/raw), aspect table, short facts
 - Persona / disk overlays merge via DeepMerge (matching keys overwrite)
-- Always injected into the system prompt; UI fills **empty** SwarmUI fields from Exact (no LLM call)
+- Always injected into the system prompt; UI fills **empty** SwarmUI fields from Exact (no LLM call). On **Generate**, the client also **overwrites** non-empty leftovers for `steps` / `cfg` / `sigma_shift` with Exact turbo|raw for the live checkpoint (unless the user asked for different numbers this turn).
 - Chat-session overrides (`session_exact`) last until persona change or clear chat — not written to disk
 - Priority: core → user message → **About the user** → session_exact → exact (+ persona) → live fields → craft `memory_hits`
 
@@ -124,8 +128,9 @@ Two layers in `memory/assistent.sqlite` (`persona` column; empty = shared):
 - `settings.json` and persona overlays stay files (layered merge + git). `.assistent.json` cards stay next to weights.
 ## VRAM handover
 
-- Before Generate the chat model is unloaded only if **Park LLM** is enabled (`keep_alive: 0`) — default off
-- After Generate the chat model is **always** force-warmed (`AssistentWarmLlm`) — Krea still often evicts VL from VRAM even without park
+- Before Generate the chat model is unloaded only if **Park LLM** is enabled (`keep_alive: 0`) — default **off** (VL reload is often 30–120s; enable only when Generate OOMs)
+- Park is a no-op when Ollama `/api/ps` already shows the chat model unloaded
+- After Generate the chat model is force-warmed (`AssistentWarmLlm`, `keep_alive: 15m`) so the next chat turn is not a surprise cold load — **but** warm is a no-op when `/api/ps` already lists the model (skips the ~20–30s reload when Krea did not evict VL)
 - Embed / memory models are never parked — reloading them would stall every retrieve
 
 ## UX