fix(iceberg): close the gaps the post-merge inspection of #707 found

[fas89]

Description

This fixes the P2 findings from the post-merge inspection of #707 (one catalog-kind table for every emitter). The inspection was a world-class review across docs, codebase/security, tests and architecture, with every finding checked by an adversarial verifier and the CLI UX walked by hand. All findings below were reproduced first and are covered by tests that fail without the fix.

Finding	Before	After
CLI output loses [...]	Rich parsed messages as markup. fluid validate printed exposes declares governance.lakeFormation, bundle findings lost [error], and pip install 'pkg[gcp]' lost its extra.	Validation lines, CLIError context and suggestions, and FluidCLIError.format_for_user print literally. tofu state rm <address> stays on one line.
Preflight skipped hand-written sinks	Runners preflighted only a derived config. A hand-written sink_connector_config that validate refused was still deployed.	One gate, iceberg_sink_plan, shared by the validator, the preflight and both runners. Hand-written configs are checked before anything is created.
Preflight and runner on different builds	Preflight used build_id, but the runners read builds[0] properties.	Runners read the build they execute (executing_build).
GCP absent catalog split	The sink wrote through REST while dbt-bigquery and the GCP IaC created BigLake, and validate was silent.	Validate error, and run error, unless the catalog type that reaches the worker is BigLake too.
catalog: bigquery on Kafka Connect	Emitted type=bigquery, which sink 1.9.2 cannot load. Silent.	Warning. The Nessie and BigQuery warnings follow the type that reaches the worker.
Catalog-move guard false positive	A Glue database left by a removed parquet expose was flagged, and nothing could get past it.	Flagged only when the moved expose's own Glue table is in state.
Guard wiring untested	Only a source-text order test existed; disabling the call passed 180 tests.	Behaviour tests through apply_via_opentofu for AWS and Snowflake. Probe failures log at WARNING.
No Snowflake upgrade path	External-catalog Iceberg on Snowflake lost its EXTERNAL VOLUME, so the next apply would plan to drop it.	Per-plugin CatalogMoveSpec. Snowflake volumes are guarded with tofu state rm remediation.
Native planner accepted unknown kinds	catalog: glu planned buckets and no Glue table.	Refused like the IaC, with a typed UnknownIcebergCatalogError and no plan_failed log line. A typo no longer also draws a Lake Formation refusal.
Matrix overclaimed coverage	The docstring said "EVERY emitter".	The agreement matrix now covers the planner, the policy compiler, Confluent and dbt-BigQuery.

P3 findings, plus the out-of-scope defects the reviewers found (the meltano, dlt and airbyte runners also read builds[0], among others), are filed on Trello.

Type of Change

• Bug fix (non-breaking change that fixes an issue)
• New feature (non-breaking change that adds functionality)
• Breaking change (fix or feature that would break existing functionality)
• Documentation update
• Refactor / code cleanup
• New provider or provider enhancement

The breaking change: a GCP Iceberg expose that a streaming sink writes to must now name its catalog. It is called out in the CHANGELOG upgrade notes, with the remedy.

Documentation

• Docs PR linked:
  • Docs PR: docs(iceberg): document the forge-cli catalog follow-up fixes (fix/iceberg-catalog-followups) forge_docs#145 (CLI quotes regenerated with exposes[<id>], GCP error, bigquery warning, Snowflake guard and upgrade path, exact preflight scope, Snowflake ownership sentence narrowed, troubleshooting rows)
• No docs needed
• Docs TODO

Checklist

• I have read the Contributing Guide
• My code follows the project's coding standards
• I have added tests that prove my fix/feature works
• All new and existing tests pass (pytest) — 4 local-environment failures fail identically on main, see Testing
• I have run ruff check and black (24.10.0) with no errors
• New maintained Python files include license headers

Testing

• Full suite (-n 8 --dist loadscope): 20905 passed, 556 skipped. Four tests fail locally, and they fail identically on main (9acc511d):
  • tests/iac/test_iac_moto_e2e.py (×3): Glue catalog id under OpenTofu 1.12 and a newer moto. A separate fix is in progress.
  • tests/test_coding_agent_live.py: local claude CLI; not run in CI.
• Lint: ruff, black --check (24.10.0) and scripts/mypy_strict.py are clean.
• Reviews:
  • Each slice was reviewed independently, with the finding's repro re-run plus a mutation.
  • A cross-slice seam review then found 5 boundary defects, all fixed here with tests:
    1. a GCP error on hand-written configs;
    2. the bigquery warning on hand-written configs;
    3. a contradictory disagreement error on GCP;
    4. a duplicated plan_failed line;
    5. a cascading Lake Formation refusal on a typo.
• Security review of 9acc511d..b87fd3a7: no findings.
  • Redaction still runs on every printed line.
  • The preflight never echoes config values.
  • The unknown-kind governance skip is backed by the validate error and by the IaC refusal at apply.

Fixed before merge: defect scan of this PR

A defect scan of this branch (41 agents across static, dynamic and CLI-journey lanes, each finding checked by a skeptic) found four gaps in this PR's own checks. All four are fixed here, each with a reproduce-first test and a mutation-killed regression test:

1. Table-less expose (a regression): the catalog-move guard planned to destroy the v0.19.0 Glue database of a moved Iceberg expose that names no table. A database is now flagged when the expose's own table is in state, or when the expose names no table.
2. Snowflake guard after a storage change: the guard missed the old volume when the upgrade also changed location.warehouse or dropped iam_role_arn. It now finds the volume by its contract-derived name, which a test pins to the emitter's own key.
3. GCP gate through catalog-impl: the gate missed sinks that reach their catalog through catalog-impl, such as sink.catalog: glue or a hand-written impl. It now keys on the catalog that reaches the worker.
4. Hand-written config selecting a different catalog: a hand-written or override type/catalog-impl that selects a different catalog than the expose's is now an error on every platform. Following the GCP remedy while a hand-written config wrote REST used to pass. Glue's Iceberg REST endpoint users declare catalog: rest.

The catalog impl class names were checked against apache/iceberg CatalogUtil.

The other 21 confirmed defects, including one pre-existing High (a streaming sink always writes the first Iceberg expose), are filed on Trello.

Full suite after the fixes: 20947 passed, 555 skipped. The same 4 local-environment failures fail identically on main.

Tested live

The real fluid CLI built from this branch:

$ fluid validate A_typo.fluid.yaml        # location.catalog: lakekeper + governance.lakeFormation
 1. expose 'orders': binding.location.catalog 'lakekeper' is not a catalog kind FLUID knows. Use one of: ...
   (one error; no Lake Formation refusal naming the typo as a catalog)

$ fluid validate B_lk_with_lf.fluid.yaml  # catalog: lakekeeper + governance.lakeFormation
 1. exposes[orders] declares governance.lakeFormation, but its Iceberg table lives in the 'lakekeeper' catalog ...

$ fluid validate gcp_absent.yaml          # GCP Iceberg sink target, no location.catalog
 1. iceberg sink (build 'ingest_events'): the GCP Iceberg expose sets no binding.location.catalog, so it is read two ways: the sink would write through a REST catalog (the 'gcp' platform default) while dbt-bigquery and the GCP IaC ... create a BigLake metastore table. Set binding.location.catalog: bigquery, or the REST kind your catalog is (e.g. rest, lakekeeper)

$ fluid generate iac A_typo.fluid.yaml --provider aws
❌ unsupported_binding  [ERR_UNSUPPORTED_BINDING]
  kind: unknown-iceberg-catalog
  error: exposes[orders] names Iceberg catalog 'lakekeper' (location.catalog), which FLUID does not know, ...
   (no plan_failed JSON line before it)