feat: validate terminal steps with unmaterialized stream output at build time

[mvallebr]

Summary

Reject at build time terminal steps whose output is a stream that will never be consumed. This is the root-cause fix for the SyncFanout._put_terminal data-loss/deadlock issue identified in #34.

The problem

A terminal step (no consumers, not hidden) whose output type is Iterator/Generator (sync) or AsyncIterator/AsyncGenerator (async) and whose output is not materialized (needs_materialize is False) produces a stream that nobody drains. At runtime this causes either:

• deadlock — the bounded handoff pump blocks forever trying to deliver the EOF marker, or
• silent data loss — the pump discards items via get_nowait() to force the EOF through

The validation

validate_no_unmaterialized_terminal_streams in dag_steps.py, called from build_dag after materialized_deps are computed. Uses the existing dag.needs_materialize() flag — no new materialization logic.

The exports step of a sub-pipeline is skipped because it will have consumers in the parent pipeline.

Bug fix: force_materialize propagation

force_materialize was not propagated through sub-pipeline expansion (dag_expansion.py). Fixed by adding force_materialize=sub_step.force_materialize to the expanded Step.

Test fixes

Existing tests with latent terminal Iterator-returning steps (the output was silently lost):

• corpus complex_parallel_mixed: step3 changed from -> Generator to -> list[int]
• observer runtime: lazy_consumer/passthrough changed from -> Iterator to -> None (drain input in body)
• dag_materializer: added force_materialize=True (tests target UNRUNNABLE error, not terminal stream)
• dag_expansion: added force_materialize=True (tests target DAG structure, not execution)
• materializers_ergonomics: added force_materialize=True on sub-pipeline gen step
• async_runner_basic: added force_materialize=True (test targets RuntimeError, not terminal stream)

New tests

6 new tests in test_dag_builder_validation.py covering:

• terminal Iterator output → raises
• terminal AsyncIterator output → raises
• terminal Iterator + force_materialize=True → builds
• terminal None output (consumer drains) → builds
• non-terminal Iterator output → builds
• exported Iterator in child pipeline → builds

Documentation

Updated DESIGN_PHILOSOPHY.md observer contract: when output is an Iterator/AsyncIterator, the observer receives it directly (via tee) and must consume it fully. An unconsumed iterator causes tee buffer growth and silently lost data. This is application responsibility.

Verification

• ruff format / ruff check --select F401: clean
• pytest tests/: 451 passed in 2.23s
• Patch coverage: 100% (14/14 trackable lines)