fix: address self-review findings across engine, rocky-sdk, and dagster

[hugocorreia90]

Self-review of the unreleased delta (the tags → Dagster fan-out, config-group guard, surrogate keys, and state-format goldens). Each finding was traced consumer-first and fixed with a discriminating test that fails without the fix.

High

• dagster: sanitize Dagster tag values, not just keys. get_model_tags sanitized governance tag keys but passed values through verbatim. Dagster validates values too (^[A-Za-z0-9_.-]{0,63}$) at AssetSpec construction, so a governance value carrying @, /, whitespace, or exceeding 63 chars raised DagsterInvalidDefinitionError when building the derived-model specs. Added _sanitize_tag_value (same charset as the key sanitizer; empty values preserved, since Dagster allows them) applied to governance and synthesized rocky/* values. New test pushes gnarly values through build_model_specs (fails on main).

Medium

• rocky-sdk: TestResult / CiResult couldn't parse real failure output. Both hand-written shadows declared failures: list[list[str]] (positional tuples), but the engine serializes failures as [{name, error}] objects — so the shadow parsed only the all-pass case and raised on any real failure, and also lacked model_results / declarative / unit_tests. Soft-swapped both names to the generated TestOutput / CiOutput, matching the existing HistoryResult / ModelHistoryResult precedent in the same module; kept the Result names as aliases for import compatibility. Real wire shape captured from rocky test --output json; discriminating round-trip tests added on both the SDK and dagster sides (the old fixtures + assertions were stale-positional, internally consistent with the broken shadow).
• engine: deny_unknown_fields on SurrogateKeySpec and TestRef. A typo'd key in a [[surrogate_key]] or [[use_test]] block was silently dropped; it now fails the load loudly.

Low / hardening

• engine: resolve_group_schema validates the resolved schema with the same validate_identifier format_table_ref applies at SQL-gen, so a bad [args] value fails at load with a clear message instead of as broken SQL (load/run stay consistent).
• engine: emit-sql guards the per-model output filename against path traversal (skip-with-report on an unsafe name) and the module doc no longer over-claims "identical to a run" for multi-adapter projects (one dialect is resolved for all models).
• engine: apply_surrogate_keys documents its validate-before-interpolation precondition and re-checks it with a debug_assert.
• engine: init_db notes the eager-table registration contract that keeps the on-disk-format golden complete.

Open decision (documented, not changed)

rocky test's model-execution path (execute_locally) does not apply declared surrogate keys, unlike rocky run and rocky emit-sql. A model with a [[surrogate_key]] is materialized without the surrogate column under rocky test, so a test depending on that column behaves differently from a real run. Left as a documented divergence rather than changed here, because aligning it would flip existing rocky test outcomes — flagging for a call on whether rocky test should mirror run-time surrogate keys.

Verification

• cargo fmt --check, cargo clippy --all-targets --all-features, cargo test (rocky-core/engine/duckdb/cli) — green.
• SDK + dagster pytest (incl. -W error::pytest.PytestCollectionWarning) and ruff check / ruff format --check — green.
• just codegen and just regen-fixtures — no drift.

🤖 Generated with Claude Code