[S2] No cross-pipeline convergence test — Backend and refactor pipelines can drift silently

[erinepshovel-code]

Two full transcript-processing pipelines exist (see #40). There is no test that runs the same transcript through both and compares outputs.

Without such a test, the drift instances catalogued in #40 (contraction tokenization, SYS/TOOL round anchoring) and any future divergences are invisible in CI. A metric value change caused by a backend bug, a core/ bug, or a spec ambiguity all look the same to the golden-drift harness: only one pipeline ran.

Proposed test structure:

# tests/test_pipeline_convergence.py

FIXTURE = "path/to/shared_fixture_transcript.txt"

def test_backend_and_core_agree_on_round_count():
    backend_result = run_backend_pipeline(FIXTURE)
    core_result    = run_core_pipeline(FIXTURE)
    assert backend_result["round_count"] == core_result["round_count"]

def test_backend_and_core_agree_on_bone_counts():
    ...

For metrics where documented divergence is intentional (e.g. contraction splitting), the test should assert the known difference explicitly rather than skipping:

# Documented: backend splits contractions, core does not.
# Backend bone_count > core bone_count by ~N per contraction in fixture.
assert backend_result["bone_count"] >= core_result["bone_count"]

This forces divergences to be named and reasoned about rather than silently tolerated.