[mvallebr]

Summary

Behavioral follow-ups from the audit, covering #128, #129, #131 and #132 (behavioral items only; #130 DagNode split and the GWT rename churn stay separate on purpose).

Important correction to #128 first: deeper investigation (watchdog + faulthandler stack dumps, 40 hot-loop + cold-process runs) showed the sync engine never deadlocked on nested_fanout — the original repro hung because my debug script drained executor.outputs after run(). What actually hangs is that footgun, and it is what this PR fixes — plus the discovery that complex_parallel* really do deadlock both engines by design (details below).

Changes

#128 — loud post-run stream semantics

• SyncQueueIterator now tracks why it closed: exhausted branches keep raising StopIteration (iterator protocol preserved), failed branches re-raise their error, and a branch closed before being fully consumed raises a descriptive RuntimeError instead of blocking forever on a dead pump
• Executor cleanup terminates live fan-outs with the new FanoutStreamClosedError (core.exceptions) — a late drain of an unconsumed branch fails loud instead of hanging or silently returning a truncated stream; AsyncQueueBranch gains identical closed-state semantics plus terminate(), and async cleanup terminal-izes tracked branches (parity)
• wait_for_workers_after_shutdown gains a log grace window (2s default): the first poll almost always sees mid-teardown workers, so the "4 workers blocked" warning fired on every healthy run (0.1s teardown) and destroyed trust in the diagnostic
• nested_fanout pack step_results corrected to the observable contract (terminal drained EACH steps publish None)

#129 — unified runtime corpus

• New tests/execution/test_corpus_runtime.py executes every registered pack on its own engine and asserts frozen step_results — replacing the two hand-maintained per-engine pack name tuples (the gap that let complex_parallel*/nested_fanout silently never run); the per-engine test_runner_step_results.py files are removed and the error-materializer registration tests moved into the unified module
• Exclusions are explicit and symmetric-checked: complex_parallel + complex_parallel_mixed deadlock BOTH engines by design — their terminal step consumes sibling fan-out branches sequentially, which circular-backpressures a bounded lockstep handoff. That consumption pattern violates the lockstep contract (§1.7); mixed_fanout shows the supported zip pattern. Fixing it would require unbounded buffering, contradicting the memory model — documented + excluded instead (kept in fix(execution): nested fan-out/fan-in topology deadlocks the sync engine #128)

#131 — fail-loud build validation (non-breaking parts)

• Include-cycle detection unified (dag_expansion.validate_no_include_cycle); the registry's instance-identity check stays separate (different semantics)
• Include return annotations validated by type inspection instead of a name substring: aliased imports now validate, and a look-alike type (NotTheParams) no longer passes; unresolvable string annotations keep the tolerant fallback
• get_safe_type_hints raises when a callable declares annotations that cannot be resolved — the previous silent {} disabled every validation that depends on hints; only fully-unannotated callables resolve to {} (PEP 563 still works — resolvable string annotations are resolved as before)

#132 — small smells

• AsyncQueueBranch.put blocks on the queue properly (wait_for-bounded, re-checking closure) instead of a 1 ms hot sleep loop
• CsvSerializer fills ragged rows with restval=""
• Async error dispatch checks callable(err_mat) before building the context, mirroring sync
• Both step runners replace the "output" not in locals() trick with an explicit _NOT_PRODUCED sentinel (provably identical semantics)

Tests

• 832 passing (was 816): +24 runtime corpus cases, +8 handoff close-semantics unit tests (sync & async), +2 hints contract tests, +2 grace-window tests, +1 corpus-symmetry check for exclusions
• Full suite, corpus spec tests, sync/async parity meta-tests, ruff format/F401 and the pre-commit coverage gate all green
• Behavioral notes for reviewers: pipelines with declared-but-unresolvable annotations now fail at build (previously silently unvalidated — this is the fail-loud contract); draining a fan-out branch post-run now raises instead of hanging