fix(VIOL-0100): wire passthrough_on_error through guard config; reject on UDF guards

[Muizzkolapo]

Summary

The passthrough_on_error guard flag was documented and already consumed by the guard evaluator and skip manager, but users could not actually set it: the guard-key allowlist rejected it at config load, and even before that the expander never forwarded it into the runtime guard dict — so the evaluator always fell back to its True default.

This wires the key end to end for SQL-style guards, on the same rails as condition/on_false:

• Allowlist (consolidated_guard.py): admit passthrough_on_error (only).
• Parser (GuardConfig): store it as an optional attribute (= True); from_dict extracts it with a bool check; from_string uses the default.
• Expander (process_guard_config): forward passthrough_on_error in the SQL branch's runtime guard dict.
• UDF guards: the UDF path always passes records through on error and has no consumer for the flag, so a non-default value on a UDF guard now raises ConfigurationError (fail loud) instead of being silently dropped.

passthrough_on_empty is intentionally left rejected — it has no runtime consumer on any path, so admitting it would deposit a dead key that changes no behavior. A test pins that it stays rejected.

Defaults are unchanged (passthrough_on_error=True); the condition/on_false flow and the semantic-error-always-bypasses rule are untouched.

Verification

• RED first: parser round-trip and expander-forward tests failed for the right reason (key rejected at the allowlist / dropped by the expander); the passthrough_on_empty rejection test passed as a baseline.
• GREEN after wiring: tests/core/utils/test_consolidated_guard.py + tests/unit/output/response/test_guard_passthrough_wire_through.py — 31 passed. New tests cover: parser round-trip, default-when-omitted, non-bool rejection, passthrough_on_empty stays rejected, UDF-guard fail-loud, expander SQL forward + default, and evaluator behavior (applies guard behavior when False, passes record when True) via the public evaluate() entry point.
• No regression: tests/unit/output/response/ (137), tests/unit/workflow/managers/ (133), tests/preprocessing/ excluding a pre-existing circular-import collection quirk (234), plus the guard/evaluator/skip/parity suites — all pass.
• ruff check clean; ruff format --check clean.

Out of scope (observed, not fixed)

• tests/preprocessing/staging/test_source_data_protection.py fails to collect on a circular import (InitialStageContext) — reproduces on clean main, unrelated to this change.
• Docs alignment for the passthrough_on_error example lives in a separate docs.agent-actions/ tree not in this checkout.