refactor phase 1: dead-code deletion + SCC rim cuts (23 → 13 modules)

[hussainsultan]

Phase 1 of the SCC-driven refactor, stacked on #295. Two workstreams: peel the strongly-connected component's periphery with mechanical edge cuts, and delete verified-dead code. The SCC ratchet from #295 enforces every cut. Full suite green (1686 passed; the count dropped from #295's 1707 by exactly the deleted dead-feature tests).

SCC rim cuts — 23 → 13 modules

• agents (and the package root) detached — agents/tools.py imported from_yaml from the package root instead of .yaml; since the root __init__ lazily imports agents, that single line kept agents.tools, agents.backends.langgraph, and boring_semantic_layer's own __init__ in the knot. One-line fix, 3 modules out.
• chart cluster detached (6 modules) — expr.py imported .chart at module level: core depending on a presentation extra, and the edge closing the cycle expr → chart → chart.utils → ops → (lazy) expr. The .chart() method now resolves the chart package at call time (pandas-.plot style); importing the core no longer loads chart at all (verified).
• convert detached — see below; deleting its dead handlers also deleted its ops import.
• flavor helpers consolidated in _xorq.py — get_ibis_module is the repo-wide ibis-flavor router (used by query, utils, ops, server, agents) and had nothing to do with nested-array compilation; it moves to the bottom layer along with _unwrap_table_proxy and null_safe_equal, and the 14-line join_utils.py is folded away. nested_compile re-exports get_ibis_module so the documented import path keeps working.

The residual 13-module SCC is the core knot (ops/expr/query/api/format/graph_utils/projection_utils/utils + serialization/*) — phase 2/3 territory.

Dead-code deletion (~1,100 lines)

• convert.py 496 → 117 lines — the 11 @convert.register handlers for ibis.expr.sql.convert were a parallel, drifted copy of the real to_untagged lowering with no caller anywhere in src or tests (the module was imported by __init__ solely "to register dispatch handlers"). Only the resolver proxies (_Resolver, _AggResolver, _PrefixProxy) survive.
• api.py 279 → 124 lines — the nine functional wrappers (join_one, filter_, group_by_, …) were pure pass-throughs to the fluent methods: never exported, never documented, used by exactly one test (rewritten fluent-style). What remains is all public API.
• legacy lambda-to-string codec deleted (~170 lines in utils.py) — expr_to_ibis_string/ibis_string_to_expr and five source-introspection helpers were the pre-v2 serialization path, superseded by the structured resolver-tree codec; only tests of the dead feature referenced them, and those are removed with it.
• graph_utils 569 → ~190 lines — nine dead exports removed (bfs, replace_nodes — real callers use the _xorq shim — to_node_safe, try_to_node, the find_* family, traverse_roots_with, extract_column_from_dimension) plus their private helpers and tests. is_field/is_table_field stay (used by build_dependency_graph).

Deliberately deferred

The three drifted twin test files (tests/ vs agents/tests/ and chart/tests/ copies of test_semantic_mcp, test_mcp_json_parsing, test_chart) are not deduped here — each pair has diverged, so merging them is a test-by-test job, not a mechanical one. The public graph_bfs family stays exported; demoting it belongs with the API-surface phase's single deprecation batch.

Verification

• python3 -m pytest src/boring_semantic_layer: 1686 passed, 1 skipped, 11 xfailed, 4 xpassed
• SCC ratchet: allowlist shrunk 23 → 13 in the same commits as the cuts
• ruff check / ruff format --check: clean
• Core import verified not to load chart; every deletion grep-verified caller-free before removal

🤖 Generated with Claude Code