fix: preserve metric semantics across fact sources

[tchivs]

Description

Fix correctness gaps in the existing Metrics prototype without introducing a separate semantic layer or serving API.

Dependency resolution

• Resolve metric dependencies iteratively, preserving forward references and shared dependencies.
• Report cycles and unknown references as ConfigError, including the dependency chain and defining file.
• Resolve metric names case-insensitively, including quoted references.

Query correctness

• Resolve and apply the complete WHERE predicate independently for each fact before aggregation. Previously, a shared filter affected only the first fact, so a filtered ratio could use an unfiltered denominator and grouped queries could return out-of-scope aggregates.
• Preserve right-only and later-source group keys across FULL JOINs, including three facts and nullable composite keys.
• Preserve computed grouping expressions, explicit dimension aliases, nested query scopes, DISTINCT ON, and named WINDOW references.
• Correct terminal reference selection without bypassing many-to-many safeguards.
• On PostgreSQL, use composite-key equality for nullable FULL JOIN keys: PostgreSQL cannot plan these joins with IS NOT DISTINCT FROM conditions.

Replace implementation-level SQL snapshots/commented scenarios with executable result/error regressions, add native PostgreSQL model-to-query integration coverage, and document the repeatable acceptance procedure.

Scope and limitations

• Metrics remains explicitly marked as a prototype, not production-ready.
• Unsupported metric WHERE subqueries, grouping sets, unresolved/ambiguous dimensions, and unsafe reference paths fail explicitly rather than silently weakening query semantics. Changing-key multi-hop reference paths are not implemented here.
• PostgreSQL group dimensions must have matching types, or be normalized with explicit casts. Metric arithmetic follows the definition dialect; fractional integer ratios and zero-denominator behavior should be expressed explicitly.
• Application authorization, a serving layer, Doris support, and additional metric types are outside this PR.

Test Plan

• make style: passed (ruff, ruff-format, mypy, migration validation).
• make fast-test: passed after installing the missing local unixODBC runtime dependency:
  • fast phase: 2664 passed, 4 skipped;
  • isolated phase: 3 passed;
  • registry-isolation phase: 1 passed;
  • dialect-isolation phase: 161 passed.
• Focused Metrics, ReferenceGraph, and native PostgreSQL suite: 54 passed.
• PostgreSQL 16.14, in a disposable isolated container: SQLMesh loads a temporary project, applies its models, compiles metric queries using Context.rewrite, and executes the generated SQL on PostgreSQL. In-memory DuckDB is used only for isolated SQLMesh state, following the existing integration fixtures.
• Native coverage includes shared filters, decimal/count ratios, empty scopes, three-fact nullable composite groups, dimension filters, conditional aggregation, leap-day/time-boundary bucketing, quoted identifiers, and DISTINCT ON.
• Native Context plan/rewrite/execute smoke verified returned results and cleanup. No business databases were used.

Checklist

• I have run make style and fixed any issues
• I have added tests for my changes (if applicable)
• All existing tests pass (make fast-test)
• My commits are signed off (git commit -s) per the DCO