[kvlonge]

Why

Warm compiles rebuilt facts that were already cached; the Polyglot fork needed upstream 0.12.1 set-operation type fixes.

Changes

• Polyglot fork 0.12.2 (includes upstream 0.12.1). Early recursive-CTE anchor binding now requires WITH RECURSIVE. When native binding fully resolves lineage but an output type is unknown, the native lineage is kept and types are recovered through the legacy path.
• SQL Rules cache unchanged non-model SQL files whose text has no macro or interpolation syntax. Files that need expansion are always re-linted.
• Cached analysis entries decode lineage into lazy indexed facts, with an eager fallback. Per-model cache digests use byte-identical orjson encoding.
• Macro modules load without deep-copying their analyzed AST.
• Discovery uses a prefix-based project-relative path helper, CSafeLoader for schema YAML, and builds path-default error context only on failure.
• The artifact writer compares stale paths as strings. SQL test artifact identities use orjson (fingerprint bumped).
• New CI guard: every fully cached fresh-process compile must take at most 0.67x the cold time (0.113.3: ~0.72, this branch: ~0.61).

Verification

• make test: 6922 passed. make check-ci passed.
• New integration tests cover:
  • mixed-type UNION ALL / UNION ALL BY NAME executed on DuckDB;
  • preserved CTE lineage under type recovery;
  • warm SQL Rules after editing a test file and a test-scoped macro.
• New unit tests cover cached lineage round-trips and digest stability.
• Paired fresh-process benchmarks (0.113.3 vs this branch) on the repository's semantic cache projects. Semantic fingerprints were identical in every run.

Models	Phase	Before	After
3000	warm	7.21s	5.78s
3000	leaf edit	5.89s	5.25s
3000	cold	9.96s	9.41s
5000	warm	11.33s	9.33s
5000	leaf edit	9.21s	8.51s
5000	cold	15.75s	14.98s
10000	warm	21.43s	16.77s