Snowflake Capability And Evidence Parity

Purpose

This spec defines how a future contractforge-snowflake adapter should map ContractForge Core contract parameters and evidence/control-table fields to Snowflake-native capabilities.

The goal is not to make the core Snowflake-aware. The goal is to give the Snowflake adapter a researched implementation target:

ContractForge contract semantics
  -> Snowflake capability declaration
  -> Snowflake SQL / task / stream / policy implementation
  -> canonical ContractForge evidence tables populated with Snowflake values

The sources for this document are Snowflake's official documentation. The adapter must verify behavior in a real Snowflake account before changing any REVIEW_REQUIRED item to SUPPORTED.

Reference Snowflake Surfaces

Snowflake surface	ContractForge use
SQL warehouse	Primary runtime for table, SQL, staged file and merge-based ingestion.
Database and schema	Mapping for target.catalog and target.schema.
Tables, views and stages	Mapping for governed table sources and staged file sources.
COPY INTO <table>	Bounded staged file loading and load validation.
Snowpipe	Continuous or near-continuous staged file loading when the contract asks for file discovery beyond simple batch loads.
Streams	Change-data capture over Snowflake tables and views.
Tasks	Scheduled execution, task graphs and stream-triggered jobs.
Row access policies	Row filter implementation.
Masking policies	Column mask implementation.
Object tags and comments	Annotation and governance metadata implementation.
Account Usage and Information Schema	Native evidence source for queries, tasks, copy loads, costs, policies, tags and object metadata.
Data Metric Functions	Optional native quality metric surface.

Primary references:

• QUERY_HISTORY
• TASK_HISTORY
• QUERY_ATTRIBUTION_HISTORY
• COPY_HISTORY
• COPY INTO table
• Snowpipe
• Streams
• CREATE TASK
• MERGE
• Information Schema COLUMNS
• Data Metric Functions
• CREATE ROW ACCESS POLICY
• Column-level security / masking policies
• CREATE MASKING POLICY
• TAG_REFERENCES
• POLICY_REFERENCES
• ACCESS_HISTORY

Adapter Subtargets

The Snowflake adapter should not declare one vague snowflake capability. It should declare explicit subtargets.

Subtarget	Status target	Purpose
snowflake_sql_warehouse	Primary	SQL execution over warehouse compute. Handles table, SQL, file-stage, quality, merge and evidence writes.
snowflake_task_graph	Primary deployment path	Renders task/task-graph deployment artifacts for scheduled project execution.
snowflake_snowpipe	Review-first	Handles staged file continuous loading where Snowpipe semantics preserve the contract.
snowflake_streams_tasks	Review-first	Handles table CDC and incremental table-to-table workflows.
snowflake_snowpark	Future	Handles complex transforms or connector behavior that is not clean in SQL alone.
snowflake_native_passthrough	Future	Delegates to Snowflake-native connectors or marketplace/native app ingestion.

Core Parameter Parity

Sources

ContractForge parameter	Snowflake mapping	Status	Notes
source.type: table	SELECT ... FROM <database>.<schema>.<table>	SUPPORTED	Adapter resolves logical refs to fully-qualified names.
source.type: view	SELECT ... FROM <database>.<schema>.<view>	SUPPORTED	Secure view behavior must be recorded in planning details when relevant.
source.type: sql	Snowflake SQL query	SUPPORTED_WITH_WARNINGS	SQL dialect is Snowflake-specific. Portable {{ table_ref:* }} placeholders remain adapter-resolved.
source.type: staged_files with format: csv/json/parquet	Named internal/external stage plus named file format or stage default file format, read through a staged query	SUPPORTED_WITH_WARNINGS	CSV positional projection, JSON VARIANT payload/projection and Parquet projection are implemented. Other formats remain review-required.
source.type: csv/json/parquet/orc/avro/xml/text	Normalize to staged_files or reviewed source-specific shorthand	REVIEW_REQUIRED	The adapter accepts explicit staged_files first so file-format, stage and evidence boundaries are visible.
source.type: object_storage	External stage over S3, Azure or GCS	SUPPORTED_WITH_WARNINGS	Storage integration, stage ownership and path privileges are environment-owned.
source.type: incremental_files	COPY_HISTORY, Snowpipe, directory table, or adapter state table	REVIEW_REQUIRED	Snowflake offers related surfaces, but exactly-once, ordering and reprocessing semantics depend on the chosen load path.
source.type: delta	External table/integration or imported stage pattern	REVIEW_REQUIRED	Delta semantics are not Snowflake-native table semantics.
source.type: iceberg_table	Snowflake-managed or external Apache Iceberg support	SUPPORTED_WITH_WARNINGS	Catalog, write support and metadata refresh behavior must be declared by the adapter.
source.type: http_file/http_csv/http_json/http_text	External access integration, Snowpark procedure, or pre-stage file path	REVIEW_REQUIRED	The adapter should prefer pre-staged files for first implementation.
source.type: jdbc/postgres/mysql/sqlserver/oracle/redshift/db2/mariadb	External function/procedure, Snowpark, or native connector pattern	REVIEW_REQUIRED	Snowflake is not a general JDBC extraction runtime by default.
source.type: kafka_bounded	Snowpipe Streaming, connector-managed stage, or external stream bridge	REVIEW_REQUIRED	Kafka replay semantics differ from ContractForge bounded Spark-style replay.
source.type: eventhubs_bounded	External bridge to Snowflake stage/Snowpipe Streaming	REVIEW_REQUIRED	Cross-cloud event offset semantics require design review.
source.type: delta_share	Marketplace/native sharing/external table pattern	SUPPORTED_WITH_WARNINGS	Needs adapter-specific source resolver.
source.type: native_passthrough	Snowflake connector/native app/marketplace source	SUPPORTED_WITH_WARNINGS	Must return explicit warnings about connector-owned semantics and evidence gaps.
source.connection_path	Core-resolved shared connection before adapter planning	SUPPORTED	Snowflake adapter receives the resolved source; it must not re-read arbitrary local paths.
source.auth	Snowflake secrets, integrations, key-pair auth or environment secret resolver	REVIEW_REQUIRED	The adapter owns credential resolution and must redact evidence.
source.watermark	SQL predicate, stream offset, COPY_HISTORY marker or state table value	SUPPORTED_WITH_WARNINGS	The implementation path changes the strength of guarantees.
source.progress_location	ContractForge state table or stage/copy marker	REVIEW_REQUIRED	The core field is portable, but Snowflake implementation must choose table state vs native metadata.
source.schema_tracking_location	Information Schema snapshot or adapter state table	REVIEW_REQUIRED	Snowflake does not need a Databricks-style schema location; the adapter should map the intent to schema evidence.

Target And Write Semantics

ContractForge parameter	Snowflake mapping	Status	Notes
target.catalog	Database	SUPPORTED	Logical catalog names should resolve through environment defaults.
target.catalog_type	database / metastore classification	SUPPORTED	Used for planning and diagnostics; not a Snowflake DDL parameter.
target.schema	Schema	SUPPORTED	Create/validate behavior is adapter-owned.
target.table	Table	SUPPORTED	Must be quoted safely by Snowflake identifier rules.
layer	Naming, comments, tags and evidence	SUPPORTED	Layer is ContractForge metadata, not a Snowflake primitive.
mode: append	INSERT INTO ... SELECT or COPY INTO append	SUPPORTED	File loads should persist COPY_HISTORY details.
mode: overwrite	Transactional replace pattern, CREATE OR REPLACE TABLE AS SELECT, or truncate/insert	SUPPORTED_WITH_WARNINGS	Atomicity and grants retention depend on the chosen pattern.
mode: upsert	MERGE INTO	SUPPORTED	Snowflake MERGE supports matched and not-matched clauses. Duplicate-source behavior must be prevalidated.
mode: hash_diff_upsert	Hash projection plus MERGE only for changed rows	SUPPORTED_WITH_WARNINGS	Hash function, null handling, collation and type casting must be adapter-defined and tested.
mode: historical	SQL MERGE/transaction pattern, possibly with streams/tasks	REVIEW_REQUIRED	Late-arriving records, effective dating and delete handling are project semantics, not a single Snowflake primitive.
mode: snapshot_reconcile_soft_delete	MERGE with matched/not matched source design	SUPPORTED_WITH_WARNINGS	Requires explicit snapshot boundary and source completeness proof.
schema_policy: strict	Compare incoming schema to INFORMATION_SCHEMA.COLUMNS and fail drift	SUPPORTED	Strict mode does not mutate target schema.
schema_policy: additive_only	ALTER TABLE ADD COLUMN after nullable additive diff	SUPPORTED_WITH_WARNINGS	Type mapping and default/nullability decisions must be explicit.
schema_policy: permissive	Reviewed DDL plan	REVIEW_REQUIRED	Type widening and incompatible changes require review.

Quality, Shape And Transform

ContractForge parameter	Snowflake mapping	Status	Notes
quality_rules.not_null	SQL COUNT_IF(column IS NULL) or DMF NULL_COUNT	SUPPORTED	Runtime SQL is the first implementation; DMFs are optional governance integration.
quality_rules.required_columns	INFORMATION_SCHEMA.COLUMNS lookup	SUPPORTED	Missing columns fail before write.
quality_rules.unique_key	SQL grouped duplicate count or DMF DUPLICATE_COUNT	SUPPORTED	Composite keys must use SQL grouping unless DMF fit is verified.
quality_rules.accepted_values	SQL NOT IN or DMF ACCEPTED_VALUES	SUPPORTED_WITH_WARNINGS	Large value sets should use reference tables instead of inline lists.
quality_rules.min_rows	SQL row count or DMF ROW_COUNT	SUPPORTED	The result goes to ctrl_ingestion_quality.
quality_rules.max_null_ratio	SQL null count divided by row count	SUPPORTED	DMF use is optional.
quality_rules.expressions	Snowflake SQL boolean predicate	SUPPORTED_WITH_WARNINGS	Expressions are dialect-specific and must be marked as such in planning.
on_quality_fail: fail	Raise error before target commit	SUPPORTED	Failure must still persist error/run evidence where possible.
on_quality_fail: warn	Persist quality evidence and continue	SUPPORTED	Run status remains independent from quality status.
on_quality_fail: quarantine	Write failed rows to ctrl_ingestion_quarantine, then write passed rows	SUPPORTED_WITH_WARNINGS	Row-level quarantine is supported for row-identifiable SQL rules; aggregate rules become recorded warnings.
shape	SQL over VARIANT, FLATTEN, casts, or Snowpark	REVIEW_REQUIRED	Nested JSON and array explosion can be mapped, but parity with Spark/Glue shape behavior must be tested.
transform.cast	Snowflake CAST / TRY_CAST depending strictness	SUPPORTED_WITH_WARNINGS	Error behavior must match the contract.
transform.derive	SQL expression projection	SUPPORTED_WITH_WARNINGS	Dialect review required.
transform.standardize	SQL string functions	SUPPORTED	Trim, upper/lower and whitespace normalization are portable enough.
transform.deduplicate	QUALIFY ROW_NUMBER() OVER (...) = 1	SUPPORTED_WITH_WARNINGS	Deterministic ordering is required.

Execution, Project And Deployment

ContractForge parameter	Snowflake mapping	Status	Notes
project.schedule.cron	CREATE TASK ... SCHEDULE = 'USING CRON <expr> <timezone>'	SUPPORTED	Snowflake task cron includes timezone in the schedule string.
project.schedule.timezone	Task cron timezone	SUPPORTED	Timezone behavior is task-specific and does not follow later session timezone changes.
project.depends_on	Task graph AFTER <task>	SUPPORTED_WITH_WARNINGS	Task graph skip/failure propagation must be captured in deployment docs.
execution.freshness	Task schedule, stream-triggered task, Snowpipe, or external scheduler	SUPPORTED_WITH_WARNINGS	Freshness is intent; enforcement depends on runtime path.
execution.latency_target	Preserved in plan/evidence and optionally task/Snowpipe design	SUPPORTED	It is not a hard runtime guarantee by itself.
execution.preferred	Scheduled, task graph, stream-triggered task, Snowpipe or external orchestration	REVIEW_REQUIRED	The adapter must block if the preferred mode cannot preserve semantics and no fallback is declared.
execution.fallback	Alternative implementation path	SUPPORTED	The planner may use the fallback only when explicitly allowed.
environment.artifacts.uri	Stage path, local output, Git artifact path, or object storage URI	SUPPORTED_WITH_WARNINGS	The Snowflake adapter should define accepted artifact schemes.
environment.evidence	Snowflake evidence database/schema	SUPPORTED	Evidence table names remain canonical ctrl_ingestion_*.
environment.parameters.snowflake.warehouse	Warehouse binding	PLATFORM_EXTENSION	Adapter-owned deployment/runtime setting.
environment.parameters.snowflake.role	Role binding	PLATFORM_EXTENSION	Adapter-owned security setting.
environment.parameters.snowflake.task_schema	Task deployment namespace	PLATFORM_EXTENSION	Keep out of ingestion contracts.

Annotations And Access

ContractForge parameter	Snowflake mapping	Status	Notes
annotations.table.description	COMMENT ON TABLE	SUPPORTED	Evidence should record applied SQL/query id.
annotations.columns.*.description	COMMENT ON COLUMN	SUPPORTED	Evidence should record per-column status.
annotations.*.tags	Snowflake object tags	SUPPORTED_WITH_WARNINGS	Tag creation, namespace and allowed values are environment/governance decisions.
annotations.*.pii	Tags and optional masking policy recommendation	SUPPORTED_WITH_WARNINGS	PII tags do not automatically enforce masking unless access contract declares it.
access.grants	GRANT statements to roles/users	SUPPORTED_WITH_WARNINGS	Privilege vocabulary needs adapter mapping. Role hierarchy may create drift.
access.row_filters	Row access policies	SUPPORTED_WITH_WARNINGS	Snowflake policies return boolean and are applied to tables/views. Principal-specific logic usually lives inside the policy expression.
access.column_masks	Masking policies	SUPPORTED_WITH_WARNINGS	Policy signature/type compatibility must be checked. Some policy combinations have platform limits.
access.access_policy.on_drift	Compare declared grants/policies with actual state	SUPPORTED_WITH_WARNINGS	Reconcile/revoke paths require explicit review due to role inheritance.
access.revoke_unmanaged	Revocation plan	REVIEW_REQUIRED	Never default to destructive access changes.

Evidence Storage Strategy

The Snowflake adapter should persist evidence in Snowflake tables with the same canonical names and columns defined by contractforge_core.evidence. These are ContractForge evidence tables, not Snowflake Account Usage views. Snowflake native views supply values; they do not replace the ContractForge evidence model.

Recommended default:

environment:
  adapter: snowflake
  evidence:
    database: CONTRACTFORGE
    schema: CF_EVIDENCE

Control tables should be append-only except lock/state helper surfaces where the adapter has a clearly documented concurrency strategy.

Control Table Mapping

ctrl_ingestion_runs

ContractForge field group	Snowflake source
Run identity	ContractForge run_id; Snowflake QUERY_HISTORY.query_id; task runs can also record TASK_HISTORY.graph_run_group_id.
Job identity	Task database/schema/name from TASK_HISTORY; manual runs use procedure/query entrypoint.
Timing	QUERY_HISTORY.start_time, QUERY_HISTORY.end_time, total_elapsed_time; task schedules use TASK_HISTORY.scheduled_time and completed_time.
Status	QUERY_HISTORY.execution_status plus adapter exception status; task status from TASK_HISTORY.state.
Error	QUERY_HISTORY.error_code and error_message; adapter exception details after redaction.
Row metrics	QUERY_HISTORY.rows_inserted, rows_updated, rows_deleted, rows_unloaded, rows_produced; adapter-computed rows_read, rows_written, rows_quarantined when native counters are insufficient.
Bytes/partition metrics	bytes_scanned, bytes_written, bytes_written_to_result, partitions_scanned, partitions_total, spill/network fields into operation_metrics_json.
Runtime	Warehouse name/size, role/user, query tag, entrypoint and CURRENT_VERSION() where collected.
Table version	Snowflake has no Delta/Iceberg-style target table version for standard tables; keep table_version_before and table_version_after null and store query/transaction markers in operation_metrics_json.

ctrl_ingestion_state

ContractForge field	Snowflake source
watermark_column, watermark_value	Contract watermark and post-read high-watermark query result.
last_success_at_utc, last_status, last_run_id	ContractForge run finalization.
last_rows_written	Adapter row metrics.
last_table_version	Null for standard tables; stream offset/copy marker may be stored in JSON.
last_watermark_candidate	Candidate max watermark before commit.
state_details_json	Stream name/offset marker, COPY_HISTORY file marker, Snowpipe pipe name or task metadata.

Snowflake Account Usage views can have latency. State that affects correctness should be written by the adapter during the run instead of relying only on Account Usage reconciliation.

ctrl_ingestion_quality

ContractForge field	Snowflake source
rule_name	Contract rule name or native DMF name.
status, severity, failed_count	Adapter SQL evaluation result.
observed_value	Count/ratio/value from SQL check or Data Metric Function.
details_json	SQL predicate, DMF name, query id, sample limits and dialect classification.

Snowflake Data Metric Functions are useful evidence inputs, but they are not a complete replacement for ContractForge runtime quality checks because ContractForge must support fail/warn/quarantine behavior during ingestion.

ctrl_ingestion_quarantine

ContractForge field	Snowflake source
record_payload	OBJECT_CONSTRUCT(*) / TO_JSON payload from failed row projection.
record_ref	Optional staging table key, target natural key or file/row reference.
reason, rule_name	Failed ContractForge row-level rule.
quarantined_at_utc	Adapter timestamp at quarantine write.

Aggregate quality failures cannot produce row-level quarantine without an explicit row predicate. The adapter should record those as warnings or failures, not invent quarantined rows.

ctrl_ingestion_errors

ContractForge field	Snowflake source
error_type, error_class	Adapter exception classifier plus Snowflake error category when available.
error_message	Redacted adapter error plus QUERY_HISTORY.error_message.
stack_trace	Procedure/Snowpark stack when available; otherwise null.
query_id / runtime_context_json	QUERY_HISTORY.query_id, role, warehouse, task metadata.

ctrl_ingestion_schema_changes

ContractForge field	Snowflake source
Before/after schema	Adapter schema snapshots from INFORMATION_SCHEMA.COLUMNS.
Applied DDL	ALTER TABLE / CREATE TABLE statement and QUERY_HISTORY.query_id.
Drift classification	Core schema policy result: strict/additive/permissive.

ctrl_ingestion_streams

ContractForge field	Snowflake source
Batch identity	Task run id, graph run group id, stream name, Snowpipe pipe name or copy batch marker.
Source progress	Stream offset state, COPY_HISTORY.file_name, last_load_time, pipe received time.
Metrics	Rows loaded, parsed rows, error count, file size, task/query metrics.
Status	Task/COPY/Snowpipe status normalized to ContractForge status.

Streaming and available-now semantics must remain REVIEW_REQUIRED until the adapter proves checkpoint/state behavior for each source family.

ctrl_ingestion_lineage

ContractForge field	Snowflake source
Input/output objects	ACCESS_HISTORY direct/base/modified objects where available.
Query identity	query_id, query tag and execution timestamp.
Event payload	Native ACCESS_HISTORY object arrays plus ContractForge source/target refs.

ACCESS_HISTORY is an audit source; the adapter should still emit a ContractForge lineage event during the run so dashboards do not depend only on Account Usage latency.

Runtime execution now writes an immediate ContractForge/OpenLineage-style COMPLETE event with the contract run id, source reference, target table, row count and Snowflake query ids. Native ACCESS_HISTORY reconciliation remains a delayed Account Usage path: it probes by structured QUERY_TAG run_id and returns PENDING without writing rows until matching native lineage rows are published.

ctrl_ingestion_annotations

ContractForge field	Snowflake source
Table/column descriptions	Applied COMMENT statements and query ids.
Tags	TAG_REFERENCES table function or Account Usage view, plus applied SQL.
Status	Applied/validated/warned/failed per annotation.

ctrl_ingestion_access

ContractForge field	Snowflake source
Grants	Applied GRANT statements and information schema/account usage checks.
Row access policies	Created/applied row access policy DDL and POLICY_REFERENCES.
Masking policies	Created/applied masking policy DDL and POLICY_REFERENCES.
Drift status	Adapter comparison of declared access contract and current Snowflake state.

ctrl_ingestion_operations

Operations fields come from ContractForge operations.yaml. Snowflake can add task name, schedule, warehouse, alert integration and query tag evidence, but it must not redefine ownership, criticality or SLA semantics.

ctrl_ingestion_explain

ContractForge field	Snowflake source
plan_text	EXPLAIN output or profile/report reference.
explain_format	TEXT or JSON when supported by the rendered statement.
captured_at_utc	Adapter timestamp.
details_json	Query id/profile URL when available.

Runtime execution captures EXPLAIN USING TEXT for the rendered write statement by default and stores failures as run metrics without failing the ingestion run. Contracts can disable capture with extensions.snowflake.explain_enabled: false.

ctrl_ingestion_cost

ContractForge field	Snowflake source
signal_name	query_count, bytes_scanned, execution_time_ms, cloud_services_credits, rows_produced, warehouse_count, attributed_compute_credits, query_acceleration_credits, later snowpipe_bytes_billed.
signal_value	QUERY_HISTORY totals, QUERY_ATTRIBUTION_HISTORY.credits_attributed_compute, warehouse metadata counts, or later COPY_HISTORY.bytes_billed for Snowpipe.
payload_json	Warehouse id/name, query id, root query id, query tag, cloud-services credits, acceleration credits, bytes scanned/written and task metadata.

Cost is an operational signal, not a billing reconciliation. Account Usage latency means cost evidence may be appended by a later reconciliation command. If Account Usage has not produced matching rows yet, reconciliation reports a pending status without inserting duplicate evidence. The same pending path is used when the executing role cannot read Account Usage, with a warning so operators can grant the needed Snowflake database role separately from runtime ingestion privileges.

ctrl_ingestion_metadata

The adapter fills ContractForge component metadata, core version, adapter version, control schema version, Snowflake account/region when configured, warehouse/role defaults and runtime feature flags. Sensitive values must be redacted.

Query Tag Requirement

The Snowflake adapter should set a structured QUERY_TAG for every statement it owns. At minimum:

{
  "product": "contractforge",
  "adapter": "snowflake",
  "run_id": "<contractforge-run-id>",
  "project": "<project-name>",
  "target": "<database.schema.table>"
}

This gives QUERY_HISTORY, QUERY_ATTRIBUTION_HISTORY, ACCESS_HISTORY and task diagnostics a stable join key back to ContractForge evidence.

Review Boundaries

The initial Snowflake planner should return REVIEW_REQUIRED or SUPPORTED_WITH_WARNINGS for these areas until proven by real tests:

• incremental_files using Snowpipe or copy history.
• Kafka/Event Hubs bounded replay.
• historical with late-arriving rows, deletes and overlapping validity windows.
• Snapshot soft delete without a proven complete source snapshot.
• Row access and masking policies with principal-specific logic.
• Tag-based masking policy generation.
• Complex shape operations over nested VARIANT payloads.
• Quarantine for aggregate quality rules.
• Type widening and incompatible schema changes.
• Cost attribution when Account Usage latency delays query-cost rows.

Initial Capability Declaration Target

The first Snowflake adapter capability set should be conservative:

Area	Initial status
Table/view/sql sources	SUPPORTED
Staged file batch load	SUPPORTED_WITH_WARNINGS
Append	SUPPORTED
Overwrite	SUPPORTED_WITH_WARNINGS
current-state upsert	SUPPORTED
hash-diff upsert	SUPPORTED_WITH_WARNINGS
historical	REVIEW_REQUIRED
Snapshot soft delete	SUPPORTED_WITH_WARNINGS
Strict schema policy	SUPPORTED
Additive schema policy	SUPPORTED_WITH_WARNINGS
Core quality rules	SUPPORTED
Quality quarantine	SUPPORTED_WITH_WARNINGS
Annotations comments	SUPPORTED
Tags	SUPPORTED_WITH_WARNINGS
Grants	SUPPORTED_WITH_WARNINGS
Row access policies	SUPPORTED_WITH_WARNINGS
Masking policies	SUPPORTED_WITH_WARNINGS
Scheduled projects	SUPPORTED
Task dependencies	SUPPORTED_WITH_WARNINGS
Cost evidence	SUPPORTED_WITH_WARNINGS
Lineage evidence	SUPPORTED_WITH_WARNINGS

Acceptance Tests For A Future Adapter

1. Plan the canonical Supabase medallion project against Snowflake and produce exact warnings/review items.
2. Render SQL for table, SQL and staged file sources without importing a Snowflake client in the core.
3. Execute append, overwrite and current-state upsert in a real Snowflake account.
4. Verify ctrl_ingestion_runs, state, quality, quarantine, errors, schema_changes, annotations, access, lineage, operations, explain and cost tables use the core schema.
5. Join ContractForge run_id to Snowflake QUERY_HISTORY through QUERY_TAG.
6. Capture task evidence for a scheduled project with at least one dependency.
7. Apply comments/tags and validate them through TAG_REFERENCES.
8. Apply row access and masking policies and validate them through POLICY_REFERENCES.
9. Prove that historical remains REVIEW_REQUIRED until the project declares the late-arriving, delete and effective-date policies needed for safe execution.