feat(benchmark): add generic evaluation and comparison harness

[wuyuanfr]

不会更改当前开发分支中 SDK/Nexent/Core 相关的内容，不更改Nexent 前后端系统服务。通过后端执行或者部署webhook服务后导入结果到Langfuse等Open Telemetry平台观测traces和实验结果。

• 主代码在sdk/benchmark/generic下，支持导入通用的agent评测集，包括不限于开源公开benchmark或者自定义数据。
• 评测指标支持 EM、F1、ACC等基础短答案QA评分器；支持缓存命中率、压缩节省tokens等多种指标观测，具体见markdown文档。

Summary

• add a generic Langfuse-backed benchmark runner with evaluator, manifest, integrity, web-evidence, and replay support
• add paired P/C and cross-worktree Legacy/P comparison runners with score synchronization and report generation
• relocate Langfuse infrastructure and ctx_debugger under sdk/benchmark
• add standard benchmark configs, bilingual documentation, and benchmark regression tests
• remove the unused production parity snapshot public entrypoint and exporter

Scope

The final diff against develop is limited to benchmark code/tests and the expected sdk/ctx_debugger relocation cleanup. It does not change sdk/nexent/core relative to the current develop branch.

Validation

• test/sdk/benchmark: 122 passed
• generic benchmark and test Ruff: passed
• ctx_debugger and LongMemEval Ruff: passed
• git diff --check: passed

Evidence

• gms8k benchmark-run screenshot

• Legacy/P comparison screenshot
  GAIA数据集实验结果显示上下文管理相比Legacy 旧版本代码提示整体提升准确率

• P/C comparison and compression screenshot

实验结果显示上下文管理提升adaptive compact策略提升缓存命中率

[codecov]

Codecov Report

❌ Patch coverage is 85.80081% with 422 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...enchmark/generic/run_context_manager_comparison.py	77.66%	55 Missing and 35 partials ⚠️
sdk/benchmark/generic/run_legacy_p_comparison.py	74.19%	34 Missing and 22 partials ⚠️
sdk/benchmark/agent_runner.py	85.15%	11 Missing and 23 partials ⚠️
sdk/benchmark/generic/tools/web_evidence.py	78.87%	16 Missing and 14 partials ⚠️
...k/benchmark/generic/evaluators/gaia_exact_match.py	74.28%	17 Missing and 10 partials ⚠️
sdk/benchmark/generic/tools/export_agent_config.py	75.28%	14 Missing and 8 partials ⚠️
sdk/benchmark/generic/run_failed_item_repeats.py	84.09%	12 Missing and 9 partials ⚠️
sdk/benchmark/generic/run_benchmark.py	95.54%	9 Missing and 11 partials ⚠️
sdk/benchmark/generic/run_integrity.py	87.09%	13 Missing and 7 partials ⚠️
sdk/benchmark/generic/common/secret_refs.py	79.41%	11 Missing and 3 partials ⚠️
... and 10 more

📢 Thoughts on this report? Let us know!