feat: zeph-bench benchmark harness

[bug-ops]

Overview

Add zeph-bench: a bench feature-gated crate that runs Zeph against standard AI-agent benchmarks (LongMemEval, LOCOMO, FRAMES, tau-bench, GAIA) in a fully automated, reproducible manner.

Spec: .local/specs/zeph-bench/spec.md

Motivation

Zeph's persistent semantic memory and tool-use capabilities have no external, reproducible measurement. This epic adds the harness to demonstrate and regression-track those differentiators against standard leaderboards.

Architecture

• New crate zeph-bench at Layer 4 (same tier as zeph-channels, zeph-tui)
• BenchmarkChannel implements zeph-core::Channel — zero changes to agent core
• Dedicated Qdrant collection prefix and SQLite DB per run — never touches production state
• bench feature flag — excluded from full bundle
• CLI: zeph bench list | download | run | show

Child Issues

• Crate scaffold and BenchmarkChannel
• CLI subcommand (zeph bench)
• Memory isolation (Qdrant + SQLite reset per scenario)
• Deterministic mode (temperature=0 override)
• LongMemEval dataset loader and evaluator
• JSON + Markdown result writer
• Baseline comparison (--baseline flag)
• Resume interrupted run (--resume flag)
• LOCOMO dataset loader
• FRAMES dataset loader
• tau-bench dataset loader
• GAIA dataset loader

Acceptance Criteria

• zeph bench run --dataset longmemeval completes end-to-end and produces valid results.json
• Two identical runs produce identical scores (determinism)
• Memory-enabled score >= memory-disabled score on LongMemEval
• No writes to production Qdrant/SQLite during bench run
• bench feature excluded from full build