Problem

QueryPlanner._can_push_op_to_sql refuses to push ORDER BY to SQL when no LIMIT follows. The intent (per the in-code comment) is to protect remote tables from full-table sorts. But the rule also fires for local from_df sources, where the data is already in the client process: there is no transport cost, and chDB's parallel ORDER BY is consistently faster than pandas single-threaded sort_values on large frames.

Result: the typical pattern ds[mask].sort_values([...]) produces a 3-segment hybrid plan (chDB filter → pandas sort), losing the chDB sort advantage.

Repro

import pandas as pd, numpy as np
from datastore import DataStore

rng = np.random.default_rng(7)
n = 10_000_000
pdf = pd.DataFrame({
    "date": ["2026-05-22"] * n,
    "code": np.array([f"{i:06d}" for i in rng.integers(1, 2001, size=n)]),
    "ix":   np.arange(n, dtype=np.int64),
    "flag": rng.integers(0, 2, size=n, dtype=np.int8),
})
ds = DataStore.from_df(pdf)
ds[ds.flag == 1].sort_values(["date", "code", "ix"]).to_pandas()

The planner logs [ORDER BY] Skipping push to SQL: no LIMIT follows, then runs the sort in pandas in segment 3.

Measured impact (chdb 4.1.8, pandas 3.0.3)

DataStore is consistently ~50% slower than the equivalent single-query chDB SQL because it skips the parallel ORDER BY.

Proposal

For an in-process PythonTableFunction source (the from_df case), push unbounded ORDER BY to SQL. Remote sources keep the current cost-aware behavior. connection.query_df already appends _row_id as a tie-breaker, so stable sort + pandas index restoration still work.

PR: #586