…n support

Adds a new REST-API ingestion mode to the Fivetran source connector and
extends destination handling to cover Fivetran's Managed Data Lake
Service (MDL) across AWS S3, Google Cloud Storage, and Azure Data Lake
Storage Gen2 — addressing customers (e.g., Angi, CUS-7715) whose Fivetran
deployment writes exclusively to a cloud lakehouse and never to Snowflake
/ BigQuery / Databricks.

== Modes ==

Two ways to read the Fivetran log:
  - `log_source: log_database` (existing) — read the Fivetran Platform
    Connector log from a Snowflake / BigQuery / Databricks warehouse.
  - `log_source: rest_api` (new) — read everything via Fivetran's REST
    API; no warehouse credentials required. Works for accounts whose
    log destination is Managed Data Lake (no SQL log to query) and for
    teams that prefer not to grant DataHub a warehouse role.

When both credential blocks are supplied the connector runs in hybrid
mode. Default is DB-primary (DB owns lineage / users / sync history;
REST fills in destination routing + Google Sheets details). Set
`log_source: rest_api` explicitly for REST-primary hybrid — REST owns
connectors / schemas / users / routing, DB log fills in per-run
DataProcessInstance events (REST has no sync-history endpoint) and
higher-fidelity column lineage.

== Managed Data Lake destinations ==

Fivetran's REST API reports `service: managed_data_lake` for every MDL
destination regardless of underlying cloud or catalog. The connector
discovers the destination via `/v1/destinations/{id}` and routes URNs
based on (in precedence order):

  1. User override on `destination_to_platform_instance.<id>.platform`.
  2. Auto-detection from MDL config toggles — currently `glue` when
     `should_maintain_tables_in_glue: true` is set on the destination.
  3. Fallback to `iceberg` (Polaris / Iceberg REST default).

Supported `platform` values for MDL destinations:
  - `iceberg` — `urn:li:dataset:(iceberg, <schema>.<table>, env)`,
    aligned with DataHub's Iceberg source over Polaris / Iceberg REST.
  - `glue`    — `urn:li:dataset:(glue, fivetran_<schema>.<table>, env)`,
    aligned with DataHub's Glue source. Auto-detected when Fivetran's
    `should_maintain_tables_in_glue` toggle is set; the `fivetran_`
    database prefix is configurable via `mdl_glue_database_prefix`.
  - `s3` / `gcs` / `abs` — path-style URNs (`<bucket-or-container>/<prefix>/<schema>/<table>`)
    aligned with DataHub's S3 / GCS / ABS sources. Bucket/prefix is
    auto-populated from the discovered destination response; user can
    override `database` to point at a different mirror layout. Note
    that DataHub's data-lake source must be configured to emit
    table-level URNs (per-folder, not per-Parquet-file) for the URNs
    to align — see the docs.

The Databricks-backed MDL variant (`service: databricks_via_managed_data_lake`)
is routed via the existing `_RELATIONAL_SERVICE_MAP` to `databricks`.

== Other improvements ==

  - Per-table column-lineage fetch from `/v1/connections/{id}/schemas/{schema}/tables/{table}/columns`
    so REST mode produces full column lineage (the bulk schemas-config
    endpoint only returns user-modified columns).
  - DB-log fallback for column lineage in REST-primary hybrid: REST
    reader prefers DB-log column-lineage rows when available (carries
    explicit `source_column_name` / `destination_column_name`) and
    falls through to per-table REST fetch otherwise.
  - Parallel per-connector REST work with bounded worker pool
    (`rest_api_max_workers`, `rest_api_per_connector_timeout_sec`),
    fail-fast pool shutdown on a single-connector hang.
  - Per-instance destination cache shared between source and log
    reader; negative cache for failed destination discovery to avoid
    re-issuing failing API calls.
  - Google Sheets connector workaround extracted to a dedicated handler.
  - REST-mode `fivetran_log_config` is optional; connector infers
    `log_source` from which credential blocks are present.

== Tests + docs ==

  - 209 fivetran tests pass (unit + integration) across the new modes
    and routing combinations.
  - New integration tests: hybrid discovery, REST-only, DB-vs-REST
    equivalence.
  - Comprehensive capability matrix and credential-coverage tables in
    `fivetran_pre.md`; recipes for `log_database` and `rest_api`.
  - Per-platform URN-shape table for MDL routing (iceberg / glue / s3
    / gcs / abs); storage-source URN-alignment caveat documented.

…laims

  - Remove the 'Migrating from managed_data_lake_destination_config'
    section: it documented a path away from a config block that this
    branch never shipped to begin with.
  - Fix the example recipe URN comment for glue (was the obsolete
    'glue.fivetran_<schema>.<table>' from the unverified prefix design;
    now 'glue.<database>.<schema>.<table>') and add the required
    'database:' field on the example.
  - Drop two stale claims that 'pinning platform skips REST destination
    discovery' — wrong after the always-on-discovery refactor; per-field
    merge already preserves user overrides while letting discovery fill
    in unpinned fields like database.
  - Update the capability matrix and credential-coverage table: REST
    mode now produces full column lineage via the per-table column
    fetch; the 'table-only by default' / 'future enhancement' framing
    is obsolete.
  - Remove a redundant REST-only example yaml inside the Hybrid mode
    section and fix the tradeoffs paragraph that misclaimed REST has a
    sync-history endpoint.

… 3.10

The per-connector timeout path catches the built-in TimeoutError, which
works on Python 3.11+ where concurrent.futures.TimeoutError is aliased
to the built-in. On Python 3.10 the two are separate classes:
`future.result(timeout=...)` raises `concurrent.futures.TimeoutError`,
which doesn't match `except TimeoutError` and bubbles out of the worker
loop.

CI failure on Py 3.10 testQuick:
TestPerConnectorTimeout::test_timeout_skips_connector_other_workers_continue
- concurrent.futures._base.TimeoutError

Catch both classes so the timeout path works across all supported
Python versions.

…ivetran_log_config

This branch lifted the three max_*_per_connector fields
(max_jobs_per_connector, max_table_lineage_per_connector,
max_column_lineage_per_connector) from FivetranLogConfig to the
top-level FivetranSourceConfig — they govern the lineage payload size
regardless of which log source (DB / REST / hybrid) the connector reads
from.

Without a compat shim, existing recipes that set those fields nested
under fivetran_log_config now fail at config-load time with
'Extra inputs are not permitted'.

Adds a model_validator(mode='before') on FivetranSourceConfig that
hoists legacy nested values to the top level with a ConfigurationWarning,
mirroring the existing compat_sources_to_database pattern. Top-level
values win on conflict.

Also adds two unit tests covering the legacy-nested and dual-spec
(top-level beats nested) cases.

This branch makes Fivetran's per-destination URN routing reflect each
destination's actual platform (via REST destination discovery) instead
of defaulting every destination to fivetran_log_config.destination_platform.

For single-destination accounts and recipes that pin every destination
explicitly, no change. For hybrid-mode multi-destination accounts the
URNs shift to correct values on next ingest, which may break dashboards
keyed on the old (conflated) URNs. Doc entry explains the change and
the opt-out path (declarative override via destination_to_platform_instance).

- Drop unused FivetranSyncHistoryResponse / FivetranSyncHistoryItem models
  (REST has no sync-history endpoint).
- Align REST get_user_email empty-string sentinel with DB reader's
  falsy check; previously every connector with no `connected_by` user
  triggered an N-group list_users walk on every cache miss.
- Count connectors_scanned only at the canonical site
  (_get_connector_workunits); REST mode was double-counting.
- Centralise Fivetran REST envelope validation in a single
  _extract_data(payload, *, context) helper; all 5 new direct
  payload["data"] accesses now route through it (consistent error
  messages, no KeyError on malformed-but-200 responses).
- Restore master's name-only connector_pattern matching in DB mode;
  the new id-OR-name match silently changed which connectors a
  pre-existing `deny: ["abc123"]` recipe would catch. REST mode
  keeps the OR-match (no compat baseline to preserve there).
- Stop synthesizing URN dataPlatform from the raw discovered.service
  string for unknown services. Previously emitted junk like
  urn:li:dataPlatform:aurora_postgres_warehouse_v2; now leaves
  platform=None so build_destination_urn raises and the existing
  per-destination dedup'd warning fires once. Regression test added.
- Add per-PlatformDetail database_lowercase: bool toggle (default True)
  with a database_for_urn @Property used at both source-URN and
  destination-URN sites. Default preserves existing lowercase URN
  behaviour; users who need to align with another DataHub source whose
  URN preserves database casing can opt out per-destination. Plain
  @Property (not @computed_field) so the derived value stays out of
  model_dump() and doesn't leak into _compose_custom_properties — the
  customProperties aspect intentionally surfaces the user-typed
  `database` verbatim. Regression test pins this.