djev (mmastrac): Jev-style decisions on DiffusionGemma

t0 = time.time()
for text, fields in CASES:
    qs = {k: {"type": "span", "instructions": q} for k, (q, _) in fields.items()}
    r = post(args.base, "/v1/systemone", {"model": "jev-latest", "state": text, "questions": qs, "seed": seed})
    for k, (q, truth) in fields.items():
        n += 1
        a = r["answers"][k]
        got = a["text"] if a["found"] else "none"
        sliced = (not a["found"]) or text[a["start"]:a["end"]] == a["text"]
        hit = norm(got) == norm(truth) and sliced
        ok += hit
        if not hit:
            print(f"  BAD {k:12s} got {got!r:28s} want {truth!r:24s} conf {a['confidence']} reads {a['reads']}{'' if sliced else '  OFFSETS WRONG'}")
print(f"spans seed {seed}: {ok}/{n} in {time.time() - t0:.1f} s")
failures += n - ok