Summary

Lifts `rocky preview create` from planning-only output to actual copy execution. For each copy-set model, resolves the source schema from the model sidecar and issues CTAS (CREATE OR REPLACE TABLE <branch_schema>.<model> AS SELECT * FROM <source_schema>.<model>) through the configured warehouse adapter.

This is the natural follow-up to PRs #279 + #280 — the planning kernel and the diff/cost math were already shipped, this one makes the prune-and-copy substrate concrete.

What's in this PR

• `execute_copy_from_base(config, models_dir, branch_schema, copy_names)` — pure builder that:
  1. Loads `rocky.toml` and builds the `AdapterRegistry`.
  2. Picks the `default` warehouse adapter (or first-declared, mirroring `rocky cost`).
  3. Loads model sidecars to resolve each copy-set model's source schema.
  4. Issues `CREATE SCHEMA IF NOT EXISTS branch_schema` once.
  5. For each model, runs CTAS via `WarehouseAdapter::execute_statement`.
  6. Per-model failures (missing-in-base, schema mismatch) surface as `copy_strategy = "failed"` without aborting.
• Registry-build failures fall back to planning-only output — the prune-set decision still ships even if the adapter is unconfigured.
• No new trait method. `WarehouseAdapter::execute_statement` already covers what we need. Introducing `clone_table_from` on the trait would block Phase 5's native-clone landing on a wider refactor; that trait method lands with Phase 5 (`SHALLOW CLONE` / zero-copy `CLONE`) using CTAS as the default impl.

What's NOT in this PR (deliberate)

• Phase 5 — adapter parity (Databricks `SHALLOW CLONE`, Snowflake zero-copy `CLONE`, BigQuery copy). CTAS works on every adapter today; Phase 5 lifts to native clones for metadata-only cost.
• Auto-invocation of `rocky run --branch ` for the prune set. Still left to the caller (Phase 4 GitHub Action orchestrates this — separate PR feat(ci): rocky preview GitHub Action + auto PR comment — Phase 4 #281).
• Phase 2.5 — checksum-bisection sampling. Independent.

Test plan

• 2 new unit tests (26 total in `preview::tests`):
  • `copy_from_base_executes_ctas_on_duckdb` — full pipeline through `AdapterRegistry` on an in-memory DuckDB. Asserts `copy_strategy="ctas"` for the happy model, `copy_strategy="failed"` for a missing-in-base model, and verifies the actual table landed in the branch schema by reading it back.
  • `copy_from_base_empty_copy_set_short_circuits` — empty copy set returns immediately without touching the adapter.
• `cargo test -p rocky-cli --lib` — 296 tests pass.
• `cargo clippy -p rocky-cli --all-targets -- -D warnings` clean.
• No schema changes — no codegen-drift risk.