fix(eval): require explicit reasoning_effort for GLM-5.3+

[ronaldtse]

Why

GLM-5.3-Flash silently defaults reasoning_effort to MAX when the
parameter is absent or unrecognized — and thinking: {"type": "disabled"} (the GLM-4.x/5.2 mechanism our eval uses) is exactly an
unrecognized value there. Our 2026-08-17 GLM-5.2 reproduction
recorded reasoning mode at minutes per long paragraph; on the
1,200-paragraph SadeedDiac-25 sweep a missing parameter would turn a
~1h run into days — silently, with no error to catch.

What changes

eval_sadeed_glm.py (the only LLM API call site across rababa /
ml-models / secryst / api — audited):

• glm-5.3* models refuse to start without an explicit effort
  value: python eval_sadeed_glm.py glm-5.3-flash low
• When given, both knobs are sent (reasoning_effort +
  thinking.type=disabled) so old and new models are covered
• Responses containing reasoning_content print a WARNING — a
  tripwire proving the disable did not take
• The effort level is part of the checkpoint filename (different
  effort = different protocol = separate resume state) and the
  startup protocol line, keeping decode-protocol disclosure intact

Verification

• py_compile clean; refusal path exercised:
  python eval_sadeed_glm.py glm-5.3-flash → exits with the message
• glm-5.2 default path unchanged (no effort → payload identical to
  the Aug 17 reproduction that scored 2.5060 raw / 2.6911 zero-skip)

Notes

• Valid effort values for the 5.3 API should be confirmed against
  z.ai docs on the first real run; an invalid value also falls back
  to MAX, which is why the response tripwire matters.
• The GLM-5.3 attempt on 2026-08-17 failed with HTTP 403 (key denied
  access) — retrying the leaderboard row is a separate decision.