Enhancing CTable, phase 2: pandas engine fix, assign()/col() chaining

[FrancescAlted]

Lands phase 2 of the CTable enhancement work (plan: plans/enhancing-ctable-phase2.md). The remaining items (first-class variable-length strings, mask-based nullable columns) are deferred to plans/enhancing-ctable-phase3.md, included here.

New features

• CTable.assign(**named_exprs): returns a view with additional computed columns — never mutates the table, never copies column data. Accepts the same forms as add_computed_column, plus Column/NullableExpr and the new unbound expressions.

• blosc2.col(name): unbound column expression. Defers operator replay until bound to a table (assign(), t[...], where()), reusing the existing Column/NullableExpr operator machinery so null propagation and comparison semantics are identical to the bound form. Enables the pandas-3 headline idiom:

  result = (
      t.assign(profit=col("revenue") - col("cost"))[col("profit") > 0]
      .sort_by("profit", ascending=False)
      .head(10)
  )

Bug fixes

• engine=blosc2.jit vs pandas 3.0.3: with the default raw=False, DataFrame.apply returned a raw ndarray instead of a properly indexed DataFrame/Series. Now reconstructs the pandas object, mirroring pandas' own raw=True path. Series.map(f, engine=blosc2.jit) is now implemented; non-numeric columns raise a clear ValueError instead of a deep numexpr error.
• CTable.head()/tail() on lazily-sorted views silently discarded sort order (they ignored _cached_live_positions); t.where(...).sort_by(...).head(10) came back in physical order. Pre-existing bug, found while testing the chain above.
• @blosc2.dsl_kernel-decorated groupby UDF aggregations crashed unconditionally (calling-convention mismatch); the wrapped plain function is now used.

Explicitly not merged

The plan's P4 (segmented reduceat acceleration for dsl_kernel groupby UDFs) was built, verified correct, and rejected by its own benchmark gate: 1.2x measured vs the required ≥5x — the groupby pipeline's partial-merging bookkeeping dominates, not the per-group UDF calls. Details in the plan's P4 implementation notes; only its crash fix (above) was kept.

Docs / examples / benchmarks

• New guide "Using Blosc2 as a pandas engine" (with the measured 4.3x axis=0 win on 1M×8, and why axis=1 is not a fit), plus bench/bench_pandas_engine.py and examples/pandas_engine.py.
• examples/ctable/assign_col.py and examples/ctable/treestore_bundle.py.
• CTable.assign/blosc2.col added to the API reference with a chained-pipelines section.

Testing

• tests/ctable (1332 tests) and tests/test_pandas_udf_engine.py all green, including new end-to-end tests through real pandas 3.0.3 and 11 new tests for col()/assign().

🤖 Generated with Claude Code