Catalog-aware relational SQL and lake query mode

[ajroetker]

Combines #141, #143, and #144 into one integration branch, then extends that base into the current catalog-aware relational SQL and lake-query architecture. The branch now covers three connected design tracks: relational typed row execution, database/namespace/tablespace catalog semantics, and external lake row sources over the same row-plan API.

What This PR Adds

Relational SQL and typed row APIs

• Lands relational mode as a table profile on the existing Antfly engine: closed schemas, typed cells, relational base-row storage, typed column entries, and derived full-text/vector/sparse/graph/algebraic indexes over that base store.
• Expands the public row-plan surface for rows/query, rows/batch, rows/aggregate, rows/window, rows/join, rows/lateral, and reusable plan envelopes with CTE metadata and result-schema reporting.
• Broadens SQL lowering into native typed plans rather than backend SQL text: DDL, row mutations, CTEs, joins, lateral plans, aggregates, windows, row claims, conflict actions, RETURNING, JSON/array/text/datetime/regex expressions, temporal DML, and schema migration metadata.
• Adds stronger relational constraint coverage: primary/unique metadata, covering payload columns, foreign-key workflows, partial/expression indexes, check/default/generated-column metadata, row-policy/auth DDL, and typed migration/rebuild intent.
• Splits the SQL adapter into parser/grammar/binder/plan modules and backs it with source fixtures plus parity fingerprints so unsupported SQL remains explicit and fail-closed.

Databases, namespaces, tablespaces, and auth

• Defines database / namespace / table as the durable catalog model, with /tables/{table} remaining shorthand for default.public.{table}.
• Adds explicit catalog route semantics and OpenAPI-backed operations for database, namespace, table, table I/O, indexes, backup/restore, documents, and tablespaces.
• Routes REST, SQL, MCP, A2A, CLI, backup/restore, foreign-source query, and role checks through typed catalog targets instead of compatibility table strings.
• Integrates qualified resources into the role model: database:name, namespace:db.ns, table:db.ns.table, and tablespace:name, preserving default-public compatibility where needed.
• Adds SQL catalog/session planning for current_database, search_path, SET/RESET/SHOW search_path, and DISCARD ALL, with session mutation modeled explicitly rather than hidden route rewriting.
• Makes tablespaces first-class metadata-backed catalog resources with SQL/OpenAPI lifecycle, durable table/database/namespace bindings, rename propagation, dependency checks, and fail-closed placement policy validation.

Lake query mode and serverless row sources

• Documents and implements the first production shape for lake tables as relational tables with an external base_source, not a separate query system.
• Adds external lake catalog bindings, pinned source snapshots, serverless manifest base sources, external inventory publication, and resolved inventory helpers.
• Builds the lake row-source path around object-storage range reads, Parquet footer/page metadata, row-group planning, typed residual filters, aggregate/query routing, and stable external row refs.
• Supports a growing Parquet scanner surface: flat numeric/text/boolean/timestamp/decimal paths, dictionary/plain pages, nullable columns, Snappy/gzip/Zstd payloads, row-group min/max pruning, object-range coalescing, and cache-keyed reads.
• Adds Iceberg metadata/inventory scaffolding through object storage: snapshot metadata reads, manifest-list/data/delete manifest planning, partition/stat pruning, cache reuse, and validation.
• Adds lake sidecar scaffolding for text, sparse, dense, graph, and algebraic materializations, with pinned source validation, candidate hydration, automatic sidecar selection, rebuild/reconcile planning, and explain/cache diagnostics.

Docs Updated

• zig/RELATIONAL.md is now the canonical relational-mode and SQL parity design, including the typed row-plan contract, constraint/migration semantics, read/write execution, lake materialization relationship, and implementation state.
• zig/DATABASES.md defines the database/namespace/table mental model, defaulting rules, explicit catalog APIs, tablespace semantics, role-resource mapping, and cross-surface integration for REST, SQL, MCP, A2A, CLI, and auth.
• zig/LAKES.md defines lake query mode as a relational row source over Parquet/Iceberg/Lance-style data, with serverless manifests, object-store reads, sidecars, cache isolation, explain/rebuild workflows, and remaining engine work.

Remaining Work Called Out In The Docs

• Finish the efficient typed ColumnBatch hot path so lake scans do not rely on JSON adaptation for broad query execution.
• Broaden Parquet/Iceberg support for nested/repeated data, page-index pruning, exact large decimals, delete semantics, schema evolution, and broader S3/GCS fixture coverage.
• Add persistent cache storage and object-store checksum provenance beyond the current in-memory range-cache validation.
• Wire lake rebuild/reconcile and snapshot-retention/GC workflows into real operator commands/controllers.
• Teach future tablespace schedulers to consume policy fields beyond the currently supported native placement keys.

Testing

• zig build
• git diff --check

Earlier focused gates on this branch also covered generated OpenAPI freshness, SQL/API parity fixtures, catalog/session SQL fixtures, explicit catalog route wiring, MCP/A2A catalog normalization, CLI surface compilation, and qualified role-resource checks.