[kvlonge]

Why

sqlbuild-python is far above the crate size limit and its helper buckets depend on each other, so it can only be split bottom-up along the XSB075 layers. This first slice moves the core every binding uses into its own crate.

Changes

• New PyO3 crate sqlbuild-python-base (about 11.8k code lines): native record, enum and error classes (specs, registration, construction, field reads); keyword calls, imports, conversions, panics; host entry calls; positional traits and macros; the allocator. Its public surface is main/ entries and role files.
• sqlbuild-python imports from the base directly, with no shims; a script rewrote the imports. It keeps the getter tables and record behaviours and passes them to register_records, so registration order is unchanged.
• python-boundary lists both binding crates; the native-layering check, XSB072/XSB073 roots and module-path scan cover both. Moved files keep their XSB072/XSB073 budgets and edges.
• The rewrite touches 14 duplicate clusters that already exist on main. The one exact copy, session_warehouse_support, is merged; the rest are listed in [dupes] with reasons.
• Caps lowered: bindings crate 217,820 → 206,319 code lines, helpers container 111 → 101 buckets.

No DuckShop case applies: only where the binding code lives changes.

Verification

• dir(sqlbuild._native) surface (types, signatures, docs, MRO, order) is unchanged.
• Dense 3,000-model compile, base and candidate alternating on one machine: compile JSON identical. Medians of 3, load 1.7–2.7 on 8 cores: wall 9.29 → 9.71 s, CPU 29.36 → 27.98 s (10 earlier runs: 9.03 → 8.76 s). Neutral.
• Incremental rebuild after a one-file edit in sqlbuild-python (--profile fast, load 5.6–8): wall 46–48 → 43–47 s, CPU 189–197 → 177–185 s.
• fensu check, check-native-layering, check-native-stub, check-build-features, clippy -D warnings, the layering, module path and fensu policy script tests, make test and make check-ci pass.