deps: split into runtime (Path 1) + training (full) requirements files

requirements-hf.txt -> requirements-runtime.txt. The old name pretended
to be HF-Space-specific; in reality it's the slim deploy set anyone
running Path 1 (workstation Ollama + Whisper) wants, regardless of
whether they're on HF Space or local. Renaming makes the README's
Path 1 install instructions honest.

requirements.txt now opens with a docstring noting it's a superset and
explicitly serving section labelled "also in requirements-runtime.txt"
so a reader doesn't accidentally pip-install both.

Dockerfile updated to COPY requirements-runtime.txt (was requirements-hf).

.gitignore: raw_record/ — DaVinci scratch dir for raw screen captures,
not for the repo.

.gitignore

@@ -35,6 +35,7 @@ env/
 
 # === Demo source captures (raw screen recordings — not for repo) ===
 *.mkv
+raw_record/
 
 # === Build artifacts ===
 llama.cpp/

Dockerfile

@@ -54,8 +54,8 @@ RUN curl -fsSL https://ollama.com/install.sh | sh
 
 # Python dependencies
 WORKDIR /app
-COPY requirements-hf.txt ./
-RUN pip install --no-cache-dir -r requirements-hf.txt
+COPY requirements-runtime.txt ./
+RUN pip install --no-cache-dir -r requirements-runtime.txt
 
 # Application code. Keep the COPY granular so the .dockerignore + the
 # requirements layer above stay cache-friendly across iterations.

requirements-hf.txt → requirements-runtime.txt

@@ -1,7 +1,7 @@
 # ============================================================================
-# Sakhi — HF Space (Docker) runtime requirements
-# Stripped vs requirements.txt: no Unsloth / PyTorch-nightly / bitsandbytes —
-# inference is via Ollama + faster-whisper (CTranslate2), neither needs PyTorch.
+# Sakhi — runtime requirements (Path 1 workstation deployment)
+# Just enough to run `python api.py`: Ollama + faster-whisper + FastAPI.
+# For training / fine-tuning, use requirements.txt (which is a superset of this).
 # ============================================================================
 
 # ASR: faster-whisper bundles ctranslate2 with its own CUDA/cuDNN libs (no system PyTorch needed)

requirements.txt

@@ -1,6 +1,10 @@
 # ============================================================================
-# MedScribe v2 — Requirements (RTX 5070 Ti, SM 12.0 Blackwell)
-# PyTorch Nightly required for CUDA 12.8 / Blackwell compute capability
+# Sakhi — comprehensive requirements (training + fine-tuning + runtime)
+# Superset of requirements-runtime.txt. Use this if you intend to retrain or
+# A/B-evaluate the Unsloth fine-tune; use requirements-runtime.txt if you only
+# need to run `python api.py` (Path 1 workstation deployment).
+#
+# PyTorch Nightly (cu128) is pinned for RTX 5070 Ti / Blackwell SM 12.0.
 # ============================================================================
 
 --extra-index-url https://download.pytorch.org/whl/nightly/cu128
@@ -25,7 +29,7 @@ datasets>=3.0.0
 # === Quantization ===
 bitsandbytes>=0.45.0
 
-# === Audio Processing ===
+# === Audio Processing (training-side) ===
 librosa>=0.10.1
 soundfile>=0.12.1
 
@@ -38,13 +42,13 @@ scipy>=1.14.0
 openai>=1.50.0
 python-dotenv>=1.0.0
 
-# === Serving ===
+# === Serving (also in requirements-runtime.txt) ===
+faster-whisper>=1.0.0,<2.0.0
+fastapi>=0.110
+uvicorn[standard]>=0.27
+python-multipart>=0.0.9
 ollama>=0.4.0
 
-# === UI ===
-# (React frontend is built separately under frontend/ via Vite; FastAPI in
-# api.py serves the dist bundle. No Python-side UI dep.)
-
 # === Utilities ===
 tqdm>=4.67.0
 pyyaml>=6.0