mdcode: generate BigQuery property-graph DDL from the Semantic Model IR

[libei]

Second in the capability-layered series merging the semantic-model tooling from the libei/knowledge-catalog fork back into upstream. Stacked on GoogleCloudPlatform#258 (Semantic Model IR + loader); it has no downstream dependencies of its own.

What this PR adds

The first destination for the Semantic Model IR: a generator that emits a single CREATE OR REPLACE PROPERTY GRAPH statement over the entities' existing base tables, with model-level metrics rendered as inline MEASURE(...) properties.

• src/libts/semantic/bigquery.ts — maps the IR to BigQuery graph DDL:

  • NODE TABLES from entities (KEY from the primary key, fields as properties) and EDGE TABLES from relationships (the association table for an M:N shape, the source base table for a direct FK), each with an explicit element KEY.
  • Measures: each model-level metric is placed on the single entity its aggregate references — a genuinely cross-table aggregate cannot be one MEASURE, so it is skipped and reported. A BigQuery graph measure may only aggregate an exposed property, so each metric is lowered into a derived operand property plus a MEASURE over it; identical operands are exposed once and reused, and COUNT(*) lowers to a count over the key property.
  • Only SUM/AVG/COUNT/MIN/MAX are emitted (the complete set BigQuery accepts inside a measure); any other aggregate is skipped + warned rather than emitted as DDL BigQuery would reject.
  • AI-first metadata (label, temporal-dimension role, description, and the aiContext instructions/synonyms/examples) is folded into OPTIONS(description=...) — the only metadata sink the graph DDL exposes — on the graph, element tables, properties, and measures.
  • Consumes dataSource as the loader's normalized, fully-qualified string; completes an under-qualified ref from options and warns if it is still short.
  • Structural drops (a keyless node, a dangling edge, an unplaceable metric) are surfaced in warnings, never silently emitted as invalid DDL.

• tests/libts/semantic/bigquery.test.ts — behavior spec over hand-built IR with inline goldens. The chain and M:N goldens are the exact DDL a live BigQuery instance accepted (validated with GRAPH_EXPAND + AGG and a GQL MATCH), plus a parsed-DDL invariant that no emitted MEASURE ever wraps a shape BigQuery rejects.

• tests/libts/semantic/bigquery.e2e.test.ts + *.bigquery.golden.sql — a file → IR → DDL golden suite over a corpus of real-shaped fixtures; the goldens are the reviewable "big picture" (open a .yaml next to its .bigquery.golden.sql). Regenerate with UPDATE_GOLDENS=1.

Scope

Property-graph DDL generation only. The CLI push path and Knowledge Catalog emit/pull follow in later PRs in the series.

Testing

• npx tsc --noEmit — clean
• bun test tests/libts/semantic/ — 109 pass / 0 fail

Review note: this draft is stacked on PR1's branch (upstream-pr1-semantic-ir-loader) within the fork, so the diff shows exactly PR2's additions. When PR1 lands upstream, PR2 will be reopened/retargeted against GoogleCloudPlatform:main.

[libei]

cc @dmitrylychagin @amirhormati — PR2 of the semantic-model upstreaming series (BigQuery property-graph DDL generator). Stacked on PR1 (upstream-pr1-semantic-ir-loader); this diff is PR2-only (12 files). GitHub wouldn't let me add you as formal reviewers on the fork since you're not collaborators here — flagging for review.