v0.19: TrainedVerifier (RouterArena 0.7358 snapshot)

[doramirdor]

Summary

Ships the trained DeBERTa-v3-small cross-encoder verifier as an MIT-licensed snapshot, fulfilling the public reproducibility commitment made in RouterArena PR #112:

The exact trained snapshot used here will be released under MIT as `NadirRouter/cascade-verifier-v1` in NadirClaw v0.14 within 14 days of this PR opening, so any reviewer can reproduce 0.7358 end-to-end.

(Numbered v0.19 because NadirClaw shipped 0.14 through 0.18 in the interim; same content, same intent.)

What lands

• `nadirclaw/trained_verifier.py` — `TrainedVerifier` class. Same `score(prompt, cheap_answer, expect_json=...)` signature and return shape as `HeuristicVerifier`. Lazy load on first call, optional preload, CPU/CUDA/MPS via `device=` constructor arg.
• `tier_config/profiles/n2_trained.yaml` — N=2 profile that routes verifier decisions through the trained model. `n2_default` stays on the heuristic, so users without the transformer stack keep working unchanged.
• `CascadeConfig` schema — new `verifier` ("heuristic" | "trained") and `verifier_model` fields. Validated; typos fail fast. Defaults preserve v0.18 behaviour.
• `NTierCascade.init` — auto-instantiates the trained verifier when the loaded profile specifies it. Lazy import so the heuristic-only path never touches torch.
• Optional extras: `pip install nadirclaw[trained]` pulls `transformers>=4.40` and `torch>=2.0`. Heuristic-only install stays zero-extra-deps.
• README: new "Trained verifier" section under "N-tier routing" — install, activation, what is and is not released.
• Tests: 9 new tests under `tests/test_trained_verifier.py`. Full suite 773/773 passing locally.

Weights release

The HuggingFace upload to `NadirRouter/cascade-verifier-v1` is pending — the local environment running this PR did not have HuggingFace auth, so I am landing the code path now and will push the weights manually right after. The local weights are at `getnadir.dev/verifier/weights/best/` (DeBERTa-v3-small, 6 layers, INT8 quantised `verifier_int8.pt` plus full FP32 `model.safetensors` and tokenizer).

Once the HF repo is live the `TrainedVerifier` constructor reaches it transparently — no code change needed.

Backward compatibility

• `n2_default` keeps the heuristic verifier. Users who do not opt into the trained extras see zero behavioural change.
• Existing `Cascade` two-tier API is untouched.
• `pip install nadirclaw` (no extras) still installs with no new dependencies.

What's NOT released

Per the RouterArena commitment we are releasing the frozen snapshot, not the pipeline:

	OSS (this PR)	Pro
Frozen verifier weights (v1)	YES (MIT)	YES
Training pipeline + corpus	NO	YES
Adaptive retraining loop	NO	YES
Custom-routed quality scoring	NO	YES

Test plan

• `python -m pytest tests/ -q` — 773 passed locally
• `n2_trained` profile parses and resolves to the trained verifier
• `n2_default` still resolves to the heuristic (backward compat)
• Empty cheap_answer short-circuits to 0.0 without invoking the model
• Interface parity with `HeuristicVerifier` (same kwargs, same return shape)
• Schema rejects unknown verifier names (typo guard)
• CI green on Python 3.10 / 3.11 / 3.12 (will verify in this PR)
• Manual smoke: load real weights from HF once published, verify a single `score()` call returns a float in [0, 1]

Repo: https://github.com/NadirRouter/NadirClaw
Service: https://getnadir.com