Make the DiariZen adapter runnable again, and pin the streaming operating point

[PhilflowIO]

Two defects found while running the next measurement on master. Both are small; the first makes a registered model impossible to run.

1. The DiariZen adapter cannot be built

The language kwarg was added to the runner's adapter dispatch (PR #16) so a hosted diarizer cannot be sent a corpus in the wrong language. The DiariZen adapter (PR #20) landed in parallel and never grew it. On master the model is registered, tested, documented — and raises TypeError on the first call:

TypeError: DiariZenDiarizer.__init__() got an unexpected keyword argument 'language'

Nothing caught it because every test builds an adapter directly, and the dispatcher is the only caller that passes all four kwargs. The adapter now accepts and ignores it, like the other local models: it scores speaker turns, not words.

The test that would have caught it reads the dispatcher's keyword names out of reproduce.py with ast and asserts every registered adapter can bind them. Parsed rather than listed, so adding a kwarg to the dispatcher tightens the contract instead of leaving a stale list that still passes. Verified against the broken signature before the fix — it fails there and passes here.

2. The streaming operating point was not in the public contract

benchmark.config.yaml documents what every published number was produced under. It pinned the collar, the overlap rule, the aggregation and the speaker mapping — and left out the one convention that moves this row most: a streaming diarizer has one DER per latency setting, and the same checkpoint reads 8.98 at 30.4 s of input-buffer latency against 9.82 at 1.04 s. That is 0.84 pp, seventeen times the ±0.05 pp reproduction tolerance, and it is the entire explanation of our disagreement with the ETH benchmark (#5454).

It is pinned now under der.model_configuration, and asserted against the adapter rather than merely for presence, so the contract file and the code cannot drift apart. A new streaming checkpoint cannot slip in unpinned either — the test sweeps every model that has a configurable operating point.

Verification

• uv run --extra dev --extra asr pytest -q → 330 passed, 2 skipped (2 new).
• uv run python scripts/verify.py → OK: 22 row(s) reproduced.
• The DiariZen AMI run that surfaced this is now proceeding on the AI box against this branch.

Refs #5454, #5461 (Forgejo Philflow/flow.raven)

🤖 Generated with Claude Code