Add lake-native Parquet and Iceberg serving foundations

[ajroetker]

Summary

• add the storage-independent row-source contract and local/external adapters
• port the complete lake-native design from the current mega-branch state in zig/LAKES.md
• add snapshot-pinned Parquet and Iceberg inventory, metadata, footer, page, row-group, predicate, delete-file, and row decoding
• add range coalescing, bounded cache behavior, object checksum validation, configured filesystem/S3/GCS credentials, and fail-closed snapshot validation
• persist external base sources and artifact references in manifests and resolve/pin external publication plans through the catalog/runtime stack
• add row-fragment publication plus text, sparse, vector, graph, and algebraic sidecar build/selection/rebuild/promotion/GC paths
• add explain/compatibility surfaces and a focused lake-test target covering the complete standalone lake stack

Porting approach

This is a surgical state port, not a cherry-pick series. Every lake-owned implementation and design file comes from the current origin/combine-pr-141-143-144 tree. Shared files were merged onto current main so the PR retains the object-store hardening already merged through #504 and does not regress newer serverless/runtime work.

The lake-owned files match the mega branch state except for compatibility with #504's expanded object-checksum contract and the review hardening listed below. Public typed relational HTTP/SQL routing remains outside this standalone PR because that adapter depends on the broader relational query and catalog surfaces intentionally excluded from the minimal #502 boundary. The complete standalone lake engine, publication stack, and row-source bridge are present here; zig/LAKES.md now marks mega-only public routing claims explicitly.

Review hardening

• publish the initial external manifest atomically even when a read-only external table has no WAL records
• publish changed .current snapshots without document-segment fallbacks, while avoiding duplicate versions for unchanged inventories
• fail closed when a .current binding has no configured discovery resolver
• validate GCS CRC32C and supported S3 full-object checksums without heap allocation, and safely ignore composite or not-yet-portable optional checksums
• keep empty external manifests honest by publishing no document lineage or search-source claims before sidecars exist

Validation

• zig build lake-test -fincremental --summary all — 275 passed
• zig build serverless-test -fincremental --summary all — 60 passed, 2 opt-in cloud tests skipped
• make zig-generated-check
• zig fmt on all changed Zig sources
• git diff --check

Depends on the object-store semantics merged in #504. The original row-source layer remains the L1 foundation in the first commit of this PR.

[ajroetker]

Scope update: the unrelated index repair-lifecycle change previously in commit 34249bf was removed from this branch. Its production-hardened replacement, including bounded cleanup, corrupt-metadata handling, restart coverage, and the stateful delete/recreate regression, is now isolated in #532.