Summary

• Adds a non-breaking WarehouseAdapter::execute_statement_with_stats trait method with a default impl delegating to execute_statement — other adapters pick up all-None stats for free.
• BigQuery overrides it to parse statistics.query.totalBytesBilled out of the REST response; those bytes flow into new MaterializationOutput.bytes_scanned / .bytes_written fields and into persisted ModelExecution, so both live rocky run and historical rocky cost produce real dollar figures on BigQuery.
• Wires the new stats through the two materialize call sites that actually build a MaterializationOutput: run_one_partition (time_interval, accumulates across the 4-statement BQ insert-overwrite transaction) and process_table (replication).
• BigQuery-only slice of Trust-system Arc 2 wave 3; Databricks / Snowflake adapters still inherit the default no-op stats and keep their existing duration-based cost formula — their slices ship as separate PRs.

Why bytes_scanned holds totalBytesBilled (not totalBytesProcessed)

BigQuery emits both. compute_observed_cost_usd's BQ branch multiplies bytes_scanned by the per-TB rate. Using processed here would under-report cost for any query under the 10 MB minimum floor Google actually charges against. The semantic impurity (field name says "scanned", holds "billed") is deliberate so downstream consumers keep treating bytes_scanned as the single cost input without BigQuery-specific branching. Documented inline at stats_from_response in connector.rs.

Scope notes (deliberately out of scope)

• Databricks / Snowflake bytes plumbing: separate PRs. Those adapters inherit the default trait method → ExecutionStats::default() → None everywhere, which keeps the existing duration-based cost formula intact.
• Plain transformation models (non-time_interval, non-replication — the execute_models loop at run.rs:2582-2611) don't construct a MaterializationOutput today. BigQuery cost attribution for derived models via full_refresh / incremental / merge still reports None after this PR. Pre-existing gap — can be closed by a follow-up that makes that loop push to output.materializations.
• snapshot_scd2 path: uses execute_query (multi-statement), stats not threaded. Another follow-up.

Test plan

• cargo test -p rocky-bigquery — new unit tests for BigQueryStatistics deserialization (present / absent / absent-query / unparseable / total_bytes_processed-only)
• cargo test -p rocky-cli — new unit tests proving populate_cost_summary emits real cost_usd when mat.bytes_scanned is Some on BigQuery, and stays None when bytes are missing
• cargo test --workspace — full suite green
• cargo clippy --workspace --all-targets -- -D warnings — clean
• cargo fmt --all --check — clean
• just codegen — regenerated schemas/run.schema.json, run_schema.py, run.ts deterministically; no drift against committed bindings
• just regen-fixtures — run twice, byte-stable (new fields are skip_serializing_if = Option::is_none and DuckDB returns None, so fixtures don't shift)
• uv run pytest in integrations/dagster/ — 312 passed
• npm run compile in editors/vscode/ — clean

Files touched (8)

• engine/crates/rocky-core/src/traits.rs — new ExecutionStats type + execute_statement_with_stats default trait method
• engine/crates/rocky-bigquery/src/connector.rs — BigQueryResponse parses statistics.query.totalBytesBilled; override threads bytes through
• engine/crates/rocky-cli/src/output.rs — MaterializationOutput gains bytes_scanned / bytes_written; populate_cost_summary + to_run_record wire them through
• engine/crates/rocky-cli/src/commands/run.rs — materialize call sites switched to execute_statement_with_stats, bytes accumulated across BQ transactions
• engine/crates/rocky-cli/src/commands/run_local.rs — snapshot_scd2 constructor carries explicit Nones
• schemas/run.schema.json, integrations/dagster/src/dagster_rocky/types_generated/run_schema.py, editors/vscode/src/types/generated/run.ts — regenerated by just codegen