feat(benchmarks): reproducible runner for four benchmarks, with the multi-round decider made visible

[YueAWu]

Summary

One runner for four long-term-memory benchmarks (LoCoMo, LongMemEval,
EverMemBench, SubtleMemory), and the multi-round decider's silent failure paths
made visible.

The decider previously fell back to a fixed top-ranked core on any failure and
returned HTTP 200 with a full episode list, so a run whose decider never
answered reported a plausible number and exited zero. It now logs at error,
surfaces on SearchData.degraded, and the harness sends one real completion to
the decider before the first question and refuses to start if it does not
answer. Twelve sweep arms were measured against that degraded path before this.

Also: each config now names the arm its published number came from, [answer]
and [judge] models included; --smoke no longer discards an explicit --conv;
the EverMemBench converter moved into its adapter so the release's
session-numbering rule is written down once instead of twice.

Area

• Architecture method
• Benchmark
• Use case
• Documentation
• Developer experience
• CI, build, or release

Verification

make lint                       clean
pytest tests/unit               2815 passed, 8 deselected
benchmarks/audit/check_protocol.py   graded protocol matches each harness

End to end, following benchmarks/README.md with steps 1-3 already done:
  DATASET=locomo CONV=0 bash benchmarks/reproduce.sh --smoke
  -> 100.0% (10/10), all four stages, 1 conversation

Config resolution under `env -i` for all four benchmarks; both newly named
answer models return content on a real completion.

Converting the EverMemBench release after moving the converter reproduces the
shipped evermembench.json byte for byte.