perf(extraction): integrate direct readers, workbook reuse and lazy backends

.github/workflows/pytest.yml

@@ -9,6 +9,21 @@ on:
     branches: [main, master]
 
 jobs:
+  base-install:
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest, windows-latest]
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
+      - name: Install base package without acceleration or development extras
+        run: python -m pip install .
+      - name: Verify installed extraction without pandas or SciPy
+        run: python -m benchmark.issue148_scipy_free --installed
+
   test:
     runs-on: ${{ matrix.os }}
     env:
@@ -26,8 +41,8 @@ jobs:
       - name: Install dependencies
         run: |
           python -m pip install --upgrade pip
-          pip install -e .[yaml,mcp]
-          pip install pytest pytest-cov pytest-mock
+          pip install -e .[yaml,mcp,fast]
+          pip install pytest pytest-cov pytest-mock pandas
       - name: Run tests (non-COM suite)
         if: runner.os != 'Windows'
         run: |

CHANGELOG.md

@@ -2,7 +2,32 @@
 
 All notable changes to this project are documented in this file. This changelog follows the [Keep a Changelog](https://keepachangelog.com/) format and covers changes from v0.2.70 onward.
 
-## [Unreleased]
+## [Unreleased]
+
+### Changed
+
+- Replaced production pandas cell extraction with direct openpyxl worksheet
+  iteration and lazy xlrd reading for legacy `.xls` files, preserving existing
+  cell normalization and serialized output.
+- Shared extraction-scoped openpyxl workbooks across compatible stages and
+  per-sheet table detection, opening the formula variant only when requested.
+- Isolated COM and rendering imports from light extraction while retaining
+  best-effort OOXML shapes/charts and legacy helper override behavior.
+- Moved SciPy to the optional `fast` extra (also included in `all`), retaining
+  accelerated selection when installed and the existing Python fallback otherwise.
+
+### Added
+
+- Added regression coverage for previous-reader parity, workbook resource
+  lifetime and loader counts, subprocess import boundaries, and clustering
+  equivalence with SciPy-free execution.
+- Added a reproducible comparison of SciPy, Python and sparse-set border
+  clustering, with exact output equivalence and recorded Windows timings.
+
+### Fixed
+
+- Preserved live pipeline backend overrides through lazy backend resolution.
+- Avoided materializing blank cells when reading a shared sparse worksheet.
 
 ## [0.8.2] - 2026-09-26
 

README.ja.md

@@ -60,9 +60,10 @@ pip install exstruct
 オプション依存:
 
 - YAML: `pip install pyyaml`
+- 罫線クラスタリングの高速化: `pip install exstruct[fast]`（SciPy）。未導入時は既存の Python 実装へフォールバックします。NumPy は引き続き必須依存です。
 - TOON: `pip install python-toon`
 - レンダリング（PDF/PNG）: Excel + `pip install pypdfium2 pillow`（`mode=libreoffice` では非対応）
-- まとめて導入: `pip install exstruct[yaml,toon,render]`
+- まとめて導入: `pip install exstruct[yaml,toon,render,fast]`
 
 プラットフォーム注意:
 

README.md

@@ -61,12 +61,13 @@ LLM/RAG pipelines and local automation.
 pip install exstruct
 ```
 
-Optional extras:
-
+Optional extras:
+
+- Border-clustering acceleration: `pip install exstruct[fast]` (SciPy). Base extraction uses the existing Python fallback when SciPy is unavailable; NumPy remains a core dependency.
 - YAML: `pip install pyyaml`
 - TOON: `pip install python-toon`
 - Rendering (PDF/PNG): Excel + `pip install pypdfium2 pillow` (`mode=libreoffice` is not supported)
-- Install everything at once: `pip install exstruct[yaml,toon,render]`
+- Install these extras at once: `pip install exstruct[yaml,toon,render,fast]`
 
 Platform note:
 

benchmark/baselines/issue148-2026-10-03-correctness.json

@@ -0,0 +1,124 @@
+{
+  "schema_version": 1,
+  "timestamp_utc": "2026-10-03T05:40:38.779016+00:00",
+  "source_revision": "5bcb0e47984104414956c9232bfca75b5076ddec",
+  "source_dirty": true,
+  "python": "3.11.13 (main, Jul 23 2025, 00:29:09) [MSC v.1944 64 bit (AMD64)]",
+  "platform": "Windows-10-10.0.22631-SP0",
+  "dependencies": {
+    "numpy": "2.4.6",
+    "scipy": "1.17.1",
+    "pandas": "3.0.6",
+    "openpyxl": "3.1.5"
+  },
+  "environment": {},
+  "mode": "light",
+  "warm_repeats": 0,
+  "check_only": true,
+  "cases": {
+    "sparse-small": {
+      "shape": [
+        40,
+        20
+      ],
+      "occupied": 44,
+      "grid_bytes": 800,
+      "grid_sha256": "0cd619ad2bd7fdf774c43d425d46517f184fc920f564310d1bdbf54db4b72e19",
+      "fixture_sha256": "bd45bc2a80e9d05e01ce2d92b9ed4d0044334bbb43bc0d4908e16727044ee0c8",
+      "min_size": 4,
+      "cluster": {
+        "scipy": {},
+        "python": {},
+        "candidate_sparse": {}
+      },
+      "light_extraction_serialization": {
+        "scipy": {},
+        "python": {},
+        "candidate_sparse": {}
+      },
+      "components": 2,
+      "output_sha256": "445a94b9270a585d90837a5300c1b69b991dd2dfacc39c4372ca602b3c0dfc64",
+      "output_bytes": 949,
+      "bbox_order_equal": true,
+      "serialized_output_equal": true
+    },
+    "sparse-large": {
+      "shape": [
+        2000,
+        200
+      ],
+      "occupied": 74,
+      "grid_bytes": 400000,
+      "grid_sha256": "89f650b3baefdf79bcd1f302e11b3c220e3bf418b721e25524e2abb2a4efb41f",
+      "fixture_sha256": "2a0c9f8b89ec25861017a8b6b81fb65f401972b9fa92a523b2b015325a02098e",
+      "min_size": 4,
+      "cluster": {
+        "scipy": {},
+        "python": {},
+        "candidate_sparse": {}
+      },
+      "light_extraction_serialization": {
+        "scipy": {},
+        "python": {},
+        "candidate_sparse": {}
+      },
+      "components": 2,
+      "output_sha256": "395c4470fa9b7fc1d1b718da9bed660289418ffbeafe03410cdf3424bb7b3957",
+      "output_bytes": 1649,
+      "bbox_order_equal": true,
+      "serialized_output_equal": true
+    },
+    "dense-style-heavy": {
+      "shape": [
+        300,
+        80
+      ],
+      "occupied": 24000,
+      "grid_bytes": 24000,
+      "grid_sha256": "9bf7697f1260d5f1781eb4878282d81007950425b289f00ffbea23d2b83c0ee8",
+      "fixture_sha256": "c6980bfa1f606bc3caad0c8c0e1204f57382e903fcacf479595c747c29b25072",
+      "min_size": 4,
+      "cluster": {
+        "scipy": {},
+        "python": {},
+        "candidate_sparse": {}
+      },
+      "light_extraction_serialization": {
+        "scipy": {},
+        "python": {},
+        "candidate_sparse": {}
+      },
+      "components": 1,
+      "output_sha256": "4027f972fe8832c95d84355c866b8235401a4453286019fa04a7a4880fdac503",
+      "output_bytes": 398897,
+      "bbox_order_equal": true,
+      "serialized_output_equal": true
+    },
+    "many-disconnected": {
+      "shape": [
+        300,
+        100
+      ],
+      "occupied": 4000,
+      "grid_bytes": 30000,
+      "grid_sha256": "4daa7f0eafcaa14efcdd9b8b9465d5011bbdc0925e5c6da6cf69d36087d5100f",
+      "fixture_sha256": "a73151afb9ae566ec8908d46473693a1edf6b0449a445072df21b2a021a6bac3",
+      "min_size": 4,
+      "cluster": {
+        "scipy": {},
+        "python": {},
+        "candidate_sparse": {}
+      },
+      "light_extraction_serialization": {
+        "scipy": {},
+        "python": {},
+        "candidate_sparse": {}
+      },
+      "components": 1000,
+      "output_sha256": "ea1cff42a0a4b84b8166f36bd8e88c3db4fd0329431f56c5c32820da916e7a1a",
+      "output_bytes": 81361,
+      "bbox_order_equal": true,
+      "serialized_output_equal": true
+    }
+  }
+}