Completion-gate validators: zero-write blind spot, build-green, literal deliverables, config/dialect lints #1174

The completion-gate validator lane (ALTIMATE_VALIDATORS_ENABLED) currently ships two checks, both of which key on "did this session modify dbt models". That leaves a structural blind spot and several unguarded failure classes observed in evaluation traces:

1. Zero-write sessions pass every gate trivially. A session that finishes without authoring anything satisfies both existing validators by default — yet "declared done without producing the deliverable" is the single most common failure signature in long-running non-interactive sessions.
2. No build-green requirement. A session can edit models, never build them (or build them with failures), and still finish.
3. No literal-deliverable check. When a task names required models/relations, nothing verifies those exact names were produced; self-chosen substitutes pass silently.
4. No structural lints for recurring correctness classes — incremental config inconsistencies (merge semantics without a unique key, missing is_incremental() guard where idempotency is required) and unguarded warehouse-specific SQL in projects that otherwise use target.type guards.

These are all checkable without knowing the expected answer — they assert structure, invariants, and the task's own literal contract. They belong in the harness-side validator lane rather than in prompt/skill guidance, which prior analysis found gets read, acknowledged, and then not followed under pressure.