release: v0.3.8 — LLM provider polish + CORS bypass + Gemini fixes

Transport
- Route every LLM fetch through tauri-plugin-http so the request leaves
  the app from Rust instead of the webview. Eliminates an entire class
  of "browser-hostile" third-party API bugs: Volcengine Ark's
  /api/coding/v3 (CORS drops Authorization), MiniMax quirks, any
  domestic cloud that doesn't configure CORS for browser origins.
  Cost: one IPC hop, negligible vs LLM latency.

New presets
- 阿里百炼 Coding Plan (single preset, OpenAI ↔ Anthropic wire toggle
  auto-swaps /v1 ↔ /apps/anthropic)
- 火山引擎 Ark (api/coding/v3 with the model catalog the Coding product
  line actually ships)
- 小米 MiMo (api.xiaomimimo.com/v1)

Preset consolidation
- New LlmPreset.baseUrlByMode field: one preset per vendor, the API
  mode toggle swaps both the wire and the URL for vendors that split
  their two compat surfaces across paths.
- custom-openai + custom-anthropic merged into a single `custom`
  preset; the API mode toggle is the only difference.

Gemini fixes
- parseGoogleLine concatenates every non-thought text part instead of
  reading parts[0] only — fixes silent truncation on 2.5/3.x reasoning
  models that split output across multiple parts in one SSE event.
  Also skips thought:true parts so chain-of-thought doesn't leak into
  the user-visible stream.
- encodeURIComponent the model segment in streamGenerateContent URLs
  so pasted OpenRouter-style ids (google/gemini-3-pro-preview) don't
  break the path.
- Fix HTTP 400 "Unknown name 'temperature'" — translate the new
  RequestOverrides type into generationConfig with Gemini-specific
  naming (top_p → topP, max_tokens → maxOutputTokens,
  stop → stopSequences). OpenAI / Anthropic wires keep the flat shape.
  Verified end-to-end against api.google.com with a real key: minimal
  body, full generationConfig, systemInstruction + user turn, and
  thinkingConfig all return HTTP 200 with expected text.

Model catalog refresh
- Kimi: k2.6 default, k2.5 / k2-thinking / for-coding kept; dropped
  moonshot-v1-* and preview-tag legacy that EOLs 2026-05-25.
- Zhipu: glm-4.6 default; added glm-4.5-{air,airx}, glm-zero-preview,
  glm-4v-plus; dropped glm-4.7 (that was Z.AI international).
- MiniMax Global + CN: M2.7 (default), M2.5; dropped legacy M2 / M2.1.
- Volcengine Ark: exact list the Coding endpoint accepts —
  Doubao-Seed-2.0-Code (default) / 2.0-pro / 2.0-lite / Seed-Code,
  MiniMax-M2.5, Kimi-K2.5, GLM-4.7, DeepSeek-V3.

Test suite: 404 → 416 passes (new Gemini parser + URL encoding cases,
new RequestOverrides translation across wires, generationConfig
mapping, multi-part / thought-filter behavior).

Author: nashsu

package-lock.json

@@ -1,7 +1,7 @@
 {
   "name": "llm-wiki",
   "private": true,
-  "version": "0.3.7",
+  "version": "0.3.8",
   "type": "module",
   "scripts": {
     "dev": "vite",

package.json

@@ -1,6 +1,6 @@
 [package]
 name = "llm-wiki"
-version = "0.3.7"
+version = "0.3.8"
 description = "LLM Wiki - A personal knowledge base for LLM concepts"
 authors = []
 edition = "2021"

src-tauri/Cargo.lock

@@ -1,7 +1,7 @@
 {
   "$schema": "https://schema.tauri.app/config/2",
   "productName": "LLM Wiki",
-  "version": "0.3.7",
+  "version": "0.3.8",
   "identifier": "com.llmwiki.app",
   "build": {
     "beforeDevCommand": "npm run dev",