build(deps): declare dependencies in pyproject.toml with docling and milvus extras

requirements.txt pinned torch and transformers in core although only docling
needs them, and on Linux torch pulls the CUDA 13 stack: 2.7 GB of the 3.0 GB
wheel download. Direct dependencies now live in pyproject.toml, uv.lock pins
everything, and application/requirements*.txt are exported from the lock by
scripts/export_requirements.sh (each file is the core set plus one extra).

The docling extra pins torch/torchvision/transformers itself and, on Linux,
resolves torch from the CPU-only PyTorch index (no nvidia packages). milvus
(pymilvus + milvus-lite, which pulls pyarrow) is the second extra.

application/core/optional_deps.py is the one place install hints come from;
the milvus store and the docling call sites use it so a missing extra fails
with the exact command to run.

scripts/export_requirements.sh

@@ -0,0 +1,64 @@
+#!/usr/bin/env bash
+# Regenerate the pip-facing requirements files from uv.lock.
+#
+# pyproject.toml declares the direct dependencies and the optional extras;
+# uv.lock pins everything. pip users, the Dockerfile and CI install from the
+# exported files, so run this after any change to pyproject.toml or uv.lock:
+#
+#   uv lock                      # or: uv lock --upgrade-package <name>
+#   bash scripts/export_requirements.sh
+#
+# Each exported file is a complete environment (core plus the named extra),
+# so `pip install -r application/requirements-docling.txt` on its own works,
+# and installing it on top of requirements.txt only adds the extra's packages.
+set -euo pipefail
+
+cd "$(dirname "$0")/.."
+
+UV=(uv)
+if ! uv --version 2>/dev/null | grep -qE '^uv 0\.([89]|[1-9][0-9])\.'; then
+  # `uv export` needs a current uv; run one through uvx without touching the
+  # machine's install.
+  UV=(uv tool run --from 'uv>=0.8' uv)
+fi
+
+export_file() {
+  local out="$1"; shift
+  local header="$1"; shift
+  local index_url="${INDEX_URL:-}"
+  {
+    echo "# GENERATED by scripts/export_requirements.sh from uv.lock -- do not edit."
+    echo "#"
+    # shellcheck disable=SC2001
+    echo "$header" | sed 's/^/# /'
+    echo
+    # uv export records the wheel's origin in uv.lock but writes no index
+    # directive; pip needs one to find the +cpu torch build.
+    if [ -n "$index_url" ]; then echo "--extra-index-url $index_url"; echo; fi
+    "${UV[@]}" export --frozen --no-hashes --no-dev --no-emit-project --no-header --quiet "$@"
+  } > "$out"
+  echo "wrote $out ($(grep -cE '^[A-Za-z0-9]' "$out") packages)"
+}
+
+export_file application/requirements.txt \
+"Core runtime. Optional extras live in requirements-<extra>.txt:
+  docling  DOC_PARSER_ENGINE=docling, docling OCR backend, read_document structured output
+  milvus   VECTOR_STORE=milvus"
+
+INDEX_URL=https://download.pytorch.org/whl/cpu export_file application/requirements-docling.txt \
+"Core runtime plus the docling extra: DOC_PARSER_ENGINE=docling, the docling
+OCR backend (layout-model hybrid OCR, ocrmac/rapidocr engines), .adoc/.vtt/.xml
+attachment parsing, and read_document's 'structured' output. The default
+anydoc engine needs none of this, and OCR itself does not either:
+OCR_ENABLED=true with the tesseract binary (or a DeepSeek-OCR endpoint) runs
+through application/parser/file/ocr_parser.py.
+On Linux torch comes from the CPU-only PyTorch index (no CUDA stack); a GPU
+deployment can reinstall torch from PyPI on top.
+Docker: --build-arg EXTRAS=docling" \
+  --extra docling
+
+export_file application/requirements-milvus.txt \
+"Core runtime plus the milvus extra (VECTOR_STORE=milvus): pymilvus and the
+embedded milvus-lite server, which pulls pyarrow.
+Docker: --build-arg EXTRAS=milvus" \
+  --extra milvus