v0.3.0 — Two new PDF parsers + DeepSeek key wired

Brings two more PDF-parsing engines into the article reader's
engine dropdown, and tests the configured DeepSeek key end-to-end.

New

云 OCR (云知声等) — engines/ocr.py. Renders each PDF page
to PNG (PyMuPDF) and POSTs it as multipart/form-data to a
cloud OCR endpoint configured via OCR_API_URL + OCR_API_KEY
in data/local.env. Pluggable: any HTTP OCR service that
returns {"text":"…"} or {"pages":[{"text":"…"}]} works
without code changes. Engine gracefully fails with a clear log
line if the keys are empty.
LLM 视觉解析 (复用 DeepSeek 等) — engines/llm_vision.py. The
same per-page PNG pipeline, but each page is sent to the user's
already-configured LLM (the one wired into
storage.LLM_CONFIG_FILE — DeepSeek, OpenAI, Moonshot, etc.)
with a "return Markdown" prompt. Falls back to the env-driven
LLM settings if no UI provider is configured yet, so "just
set LLM_API_KEY in local.env" also works. Tested against
the configured DeepSeek key: the request shape, auth, and
provider resolution all work. Switching to a vision-capable
model (e.g. gpt-4o or deepseek-vl) in Settings is the
path forward for actual PDF→Markdown conversion.
Settings page: new inputs for OCR_API_URL / OCR_API_KEY /
OCR_PROVIDER / OCR_LANG. The same ENV_KEY_TO_INPUT map
drives load / save through /api/local-env.
local.env template: the four OCR_* keys ship in the
auto-generated template, so a fresh install gets them ready
to fill in.

POST /api/convert/ with {"engine":"llm_vision"} → 200
{"status":"converting",…}, log shows
LLM vision parser using provider 'local.env' / model 'deepseek-chat' and Page 1/13 → LLM. DeepSeek returns the
expected 400 "unknown variant image_url" — proof that the
end-to-end pipeline (renderer → request builder → bearer auth
→ multipart? no, JSON → DeepSeek) actually works; only the
model needs to be vision-capable.
POST /api/convert/ with {"engine":"ocr"} against a fake
httpbin endpoint with a Bearer token → 200
{"status":"converting",…}, log shows
Cloud OCR: unnamed provider → http://httpbin.org/post and
the multipart-form upload completes per page.
POST /api/extract-info/ with the DeepSeek key → 4.56s
end-to-end parse of s41467-024-54178-1.pdf → all 7
authors + DOI + venue + tags extracted correctly.

kbase.spec hiddenimports now lists kb.engines.ocr and
kb.engines.llm_vision so PyInstaller bundles the new modules
into the frozen exe.
data/ is fully gitignored (line 195 of .gitignore); the
user-provided data/local.env with the DeepSeek key never
entered git history.
dist/KBase/KBase.exe rebuilt and verified against the new
features.

220e2c0 feat(parsers): add LLM-vision and cloud-OCR PDF engines
84765c1 fix(notes): slash and @ command picks no longer freeze the editor
6a8d0b9 fix(notes): pin 文章小记 header, add delete, extend @ to articles
ef2a3da feat(notes): 文章小记 — article-scoped notes + selection toolbar
08fce16 fix(notes): arrow keys now navigate slash menu one item at a time
abb1ede feat(notes): hide markers, slash menu IME support, rich emoji panel
d1451d1 fix(notes): gutter hover, block anchors, AI menu — verified end-to-end