Add provider-normalized context views

[fivetran-davefowler]

Summary

Adds information-schema-like context views. v1 deliberately extends only the surfaces INFORMATION_SCHEMA already has — AGENTS.TABLES and AGENTS.COLUMNS — rather than inventing new cross-provider object types.

• Each provider publishes a normalized AGENTS.<PROVIDER>_TABLES / AGENTS.<PROVIDER>_COLUMNS view (dbt, LookML, OSI).
• AGENTS.TABLES / AGENTS.COLUMNS take the native INFORMATION_SCHEMA view as the row spine via SELECT t.* and left join every provider view that exists by object identity, appending each provider's columns under a <provider>_ prefix.

Design

• SELECT t.* spine — inherits whatever native columns the account exposes; no hardcoded INFORMATION_SCHEMA column list (removes the earlier is_hybrid/last_ddl/etc. fragility).
• Generic identity merge — providers are discovered, not hardcoded; within a provider, rows are aggregated to one per identity to prevent fanout; prefixed columns mean providers never collide.
• No hardcoded memory counts — memory participates purely additively: when a memory provider later ships its own *_TABLES/*_COLUMNS view, its columns appear automatically with no change here. (Tracked separately; not in this PR.)
• Fail-soft — view creation runs at the end of each ingestion but a view error warns and is skipped, never failing the data load.
• Deferred: RELATIONSHIPS / METRICS / ENTITIES. Those are object types semantic providers (OSI) already model; the information-schema-faithful home for relationships is the REFERENTIAL_CONSTRAINTS / KEY_COLUMN_USAGE family (future), not a custom view.

Known limitation

INFORMATION_SCHEMA is per-database, so AGENTS.TABLES/COLUMNS cover the database that holds the AGENTS schema. Multi-database coverage (via ACCOUNT_USAGE) is noted as an open question in the proposal.

Verification

uv run python -m unittest discover -s tests   # 19 tests, OK
uv run python -m compileall -q src
git diff --check                              # clean

Note: tests assert generated-SQL structure; they do not execute against Snowflake.

🤖 Generated with Claude Code