feat(storage): surface the typed-table definition in table-detail (#621)

[padak]

What

storage table-detail now returns the table's definition — the typed-table layout
carrying primaryKeysNames and, on BigQuery, timePartitioning, rangePartitioning
and clustering. null for an untyped table.

Why

kbagent could already write that layout (storage create-table --source-table-id --time-partitioning-field --clustering-field, promoted with storage swap-tables) but
had no way to read it back. StorageService.get_table_detail() built its response from
an explicit field allowlist that dropped definition, even though
GET /v2/storage/tables/{id} returns it for typed tables with no include= parameter.
The write half of the repartition flow was supported; the verify half was not.

That matters because the table ID is unchanged whether the swap happened or not — the
layout is the only field that distinguishes a completed repartition from a failed one.
On a Keboola-managed BigQuery project where the caller holds no bigquery.jobs.create,
the Storage definition is also the only reachable view of the registered layout.

Changes

• service (services/storage_service.py::get_table_detail) — definition passes
  through verbatim; None on untyped tables, which needs no special-casing.
  Deliberately detail-only: the LIST endpoint genuinely does not return it, so
  storage tables would mean one detail request per table — a different change with a
  different cost profile (the issue calls this out as out of scope).

• command — human mode gains Partitioning: / Clustering: rows between the
  primary key and last-import lines:

  Table: out.c-my-bucket.my-table
    Rows: 6,290,737
    Primary key: id
    Partitioning: DAY on created_at
    Clustering: tenant_id, country
    Last import: 2026-08-19T17:26:18+0200
  

  Rows are emitted only when a layout exists, so Snowflake and untyped-table output is
  byte-identical. Ingestion-time partitioning (no field) prints just the type.

• REST — GET /storage/table-detail/... delegates to the service, so the route
  picks the field up with no router change. No new command, so no OPERATION_REGISTRY
  entry.

• file-size budget — commands/storage.py is past its grandfathered ceiling, so
  rather than golf the diff, the table-detail human renderer moved into a private
  commands/_storage_format.py (the same pattern as _storage_snapshots.py) together
  with the new layout formatter. Net effect on commands/storage.py is a 22-line
  shrink (2246 → 2224).

Robustness note

The formatter type-guards at every level (isinstance(..., dict) on the definition and
on each sub-object). A Storage API deploy once served definition as [] rather than an
object — the incident pinned by tests/test_storage_empty_definition.py, which broke the
legacy Go CLI's strict decoder. A test asserts the renderer returns nothing for that shape
so the crash class cannot come back through this path.

Version

Bumped to 0.87.1 with a changelog entry. Renumber freely if it should land under a
different version.

Testing

• make check green: lint, format, typecheck, skill-check, version-check,
  command-sync-check, changelog-check, error-codes, sentinel-guards, loc-check, and
  5711 tests passed.
• New tests/test_storage_table_definition.py: service passthrough (typed + untyped),
  seven formatter cases (no layout, non-dict definition, time partitioning with and
  without field, range partitioning with and without bounds, clustering), and three CLI
  cases (human output shows and orders the rows; human output unchanged without a layout;
  --json carries definition).
• Not verified against a live BigQuery project — no live-project access in this run. The
  response shape follows the issue's reported v0.86.0 output and the existing
  create-table write-side field names.

Note for the merge queue

Touches the same get_table_detail function as #628 (issue #624). Whichever lands second
needs a trivial rebase; there is no semantic overlap — #628 changes how column
descriptions are read, this one adds a field to the return dict.

Fixes #621

[padak]

Closing as a duplicate of #629, which was opened concurrently and is the stronger implementation: it also surfaces requirePartitionFilter and the partitions[] count, adds an E2E case, and updates keboola-expert.md / storage-types-workflow.md. Nothing here is worth porting over — the isinstance guarding against the "definition":[] wire shape (see tests/test_storage_empty_definition.py) is present in #629 too.

Opened by the daily issue-triage run before #629 existed; no reviewer time was spent on it.