refactor phase 3b: split ops.py — package conversion + compile strategies extracted

[hussainsultan]

Phase 3b of the SCC-driven refactor, stacked on #298. The 7,720-line ops.py monolith becomes an ops/ package: five helper zones and the three compile strategies move into focused modules, all as verbatim moves — every commit is green on the full suite and the SQL snapshots are byte-identical throughout. _core.py shrinks 7,720 → 5,369 lines.

Shape

Stage 1 — package conversion. ops.py → ops/_core.py verbatim; ops/__init__.py is the stable import surface (explicit re-exports of every name any module or test pulls from boring_semantic_layer.ops, plus a __getattr__ fallback to _core), so the split can never break an importer.

Stage 2 — helper zones out of _core, one commit each:

• _compat.py — xorq/plain-ibis backend interop (conversion, connection identity, backend rebinding)
• _tracking.py — projection-pushdown column tracking (ColumnTracker & friends)
• _reductions.py — the phase-2-deferred ReductionKind consolidation: reduction sniffers, re-aggregation op selection, count-identity fills — one module answers "how does this reduction re-aggregate from a finer pre-aggregate"
• _values.py — Dimension/Measure/CalcMeasure, _CallableWrapper, the dimension proxies, column-error formatting
• _normalize.py — join/group-by input normalization and collision-free join-name allocation

Stage 3 — compile strategies out of SemanticAggregateOp:

• _compile_preagg.py (1,288 ln) — fan-out-safe source-grain pre-aggregation, the dimension-bridge re-join helpers, calc-spec application
• _compile_deferred.py (235 ln) — aggregate-first + deferred dimension LEFT JOINs
• _compile_nest.py (86 ln) — nest= aggregate entries

Strategy modules import _core downward; the op methods dispatch through a call-time module accessor, so the import graph stays acyclic (the #298 ratchet stayed green through every commit — lazy in-package imports would have tripped it).

What stays in _core and why

The op classes are mutually referential (SemanticJoinOp.index() constructs SemanticIndexOp; the join analysis helpers isinstance-check the op types and are called from to_untagged dispatchers), so the index cluster and the join-tree analysis belt can't leave without inverting those references — that's the next seam. The # --- N. phase decomposition of the preagg pipeline is deliberately not in this PR: everything here is a verbatim move, and the decomposition is behavior-adjacent surgery that deserves its own diff (phase 3c) — now against a 1,288-line module instead of a 7,700-line one, with the snapshot corpus as its gate.

Verification

• Full suite after every stage; final: 1703 passed, exit 0
• SQL snapshots byte-identical across all 9 commits
• SCC ratchet green throughout (graph remains a DAG); ruff + format clean
• One extraction bug caught mid-flight by the soundness suites (a method captured into the wrong module still being called method-style) and fixed before commit

🤖 Generated with Claude Code