From a3946db6381ca84f2d70fa4a1e439dc12f374be8 Mon Sep 17 00:00:00 2001
From: FBumann <117816358+FBumann@users.noreply.github.com>
Date: Fri, 31 Jul 2026 19:33:45 +0200
Subject: [PATCH] =?UTF-8?q?perf(sinks):=20cols=20is=20positional=20?=
 =?UTF-8?q?=E2=80=94=20a=20row's=20index=20is=20its=20solver=20column?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Measured and rejected; kept for the record. See the PR.
---
 .../relational/engines/polars/executor.py     | 21 ++++++++++++-------
 src/lpspec/relational/sinks/tables.py         | 16 +++++++-------
 .../relational/sinks/writers/lp_file.py       |  4 ++--
 3 files changed, 23 insertions(+), 18 deletions(-)

diff --git a/src/lpspec/relational/engines/polars/executor.py b/src/lpspec/relational/engines/polars/executor.py
index 30867225d..1da7ff858 100644
--- a/src/lpspec/relational/engines/polars/executor.py
+++ b/src/lpspec/relational/engines/polars/executor.py
@@ -39,7 +39,7 @@
 
 #: The four frames a sink reads, as schemas. Stated here because the executor
 #: is what fills them and an empty model still has to have them.
-_COLS = ('col', 'lb', 'ub', 'vtype')
+_COLS = ('lb', 'ub', 'vtype')
 _OBJ = ('col', 'coeff')
 _ROWS = ('row', 'sense', 'rhs')
 _MATRIX = ('row', 'col', 'coeff')
@@ -159,12 +159,19 @@ def _build_variable(self, v: plan.VariableDeclaration) -> pl.DataFrame:
         self._blocks[v.name] = (start, labelled.height)
 
         bounded = self._q.bounds(labelled.lazy(), v)
-        cols = bounded.select(
-            pl.col('var_label').alias('col'),
-            pl.col('lb').cast(pl.Float64),
-            pl.col('ub').cast(pl.Float64),
-            pl.lit(v.variable_type, dtype=_DTYPES['vtype']).alias('vtype'),
-        ).collect(engine='streaming')
+        # A declaration owns a contiguous run of labels and the blocks are
+        # concatenated in declaration order, so a frame sorted within each
+        # block is sorted across all of them — which makes a row's *position*
+        # its solver column index, and the column redundant.
+        cols = (
+            bounded.sort('var_label')
+            .select(
+                pl.col('lb').cast(pl.Float64),
+                pl.col('ub').cast(pl.Float64),
+                pl.lit(v.variable_type, dtype=_DTYPES['vtype']).alias('vtype'),
+            )
+            .collect(engine='streaming')
+        )
 
         bad = cols.filter(pl.col('lb').is_null() | pl.col('ub').is_null()).height
         if bad:
diff --git a/src/lpspec/relational/sinks/tables.py b/src/lpspec/relational/sinks/tables.py
index 8bf0d80dd..e8d26fd0f 100644
--- a/src/lpspec/relational/sinks/tables.py
+++ b/src/lpspec/relational/sinks/tables.py
@@ -35,7 +35,7 @@
 class ModelTables:
     """The built model, as a sink sees it.
 
-    ``cols`` (col, lb, ub, vtype), ``obj`` (col, coeff), ``rows`` (row, sense,
+    ``cols`` (lb, ub, vtype), ``obj`` (col, coeff), ``rows`` (row, sense,
     rhs) and ``matrix`` in COO (row, col, coeff). The scalars are what a sink
     cannot cheaply recover; the objective constant lives outside the frames
     because it has no column to attach to.
@@ -96,15 +96,13 @@ def dense_columns(self, infinity: float) -> DenseColumns:
         import numpy as np
 
         count = self.column_count
-        at = self.cols['col'].to_numpy()
-        lb = _scattered(count, at, self.cols['lb'].to_numpy(), -infinity)
-        ub = _scattered(count, at, self.cols['ub'].to_numpy(), infinity)
-        integral = _scattered(
-            count, at, self.cols.select(pl.col('vtype') != 'continuous').to_series().to_numpy(), False
-        )
+        # `copy=True` rather than in place: these are views of the frame now,
+        # and rewriting an infinity through one would edit the built model to
+        # suit whichever solver asked last.
+        lb = np.nan_to_num(self.cols['lb'].to_numpy(), copy=True, neginf=-infinity, posinf=infinity)
+        ub = np.nan_to_num(self.cols['ub'].to_numpy(), copy=True, neginf=-infinity, posinf=infinity)
+        integral = self.cols.select(pl.col('vtype') != 'continuous').to_series().to_numpy()
         cost = _scattered(count, self.obj['col'].to_numpy(), self.obj['coeff'].to_numpy(), 0.0)
-        np.nan_to_num(lb, copy=False, neginf=-infinity, posinf=infinity)
-        np.nan_to_num(ub, copy=False, neginf=-infinity, posinf=infinity)
         return lb, ub, cost, integral
 
 
diff --git a/src/lpspec/relational/sinks/writers/lp_file.py b/src/lpspec/relational/sinks/writers/lp_file.py
index f809e2222..6fcd1edd7 100644
--- a/src/lpspec/relational/sinks/writers/lp_file.py
+++ b/src/lpspec/relational/sinks/writers/lp_file.py
@@ -59,7 +59,7 @@ def write_lp_file(model: ModelTables, path: str | Path) -> None:
     objective = model.obj.lazy().sort('col').select(_term(pl.col('coeff'), pl.col('col')))
     bounds = (
         model.cols.lazy()
-        .sort('col')
+        .with_row_index('col')
         .select(
             pl.concat_str(
                 _bound(pl.col('lb'), '-infinity').alias('lb'),
@@ -89,7 +89,7 @@ def write_lp_file(model: ModelTables, path: str | Path) -> None:
         _sink(bounds, f)
 
         for variable_type, keyword in (('binary', 'binary'), ('integer', 'general')):
-            chosen = model.cols.lazy().filter(pl.col('vtype') == variable_type).sort('col')
+            chosen = model.cols.lazy().with_row_index('col').filter(pl.col('vtype') == variable_type)
             if chosen.select(pl.len()).collect().item() == 0:
                 continue
             f.write(f'\n{keyword}\n'.encode())