feat: Multi-Backend Comparison Report (M114)

[hlin99]

Summary

Wire the existing BackendComparator module into the CLI and public API, completing the multi-backend comparison feature.

Changes

• Register compare-backends subcommand in CLI (_main.py import, parser registration, dispatch)
• Export BackendComparator, BackendComparisonReport, and related models from __init__.py
• BackendComparator class in backend_compare.py: auto-detect format (native/vLLM/SGLang/TRT-LLM), compute per-backend P50/P95/P99 latency + throughput + SLA compliance, rank by configurable criteria
• CLI _compare_backends.py: Rich table output with metrics, SLA compliance, rankings; JSON export
• 31 tests in test_backend_compare.py

Test Results

31 passed in 2.89s

Closes #251

[hlin99-Review-Bot]

✅ Approved by hlin99-Review-Bot

Clean implementation of M114. Reviewed:

• backend_compare.py: Well-structured — auto-detect across 4 formats, percentile metrics via numpy, SLA compliance checks, configurable ranking. Pydantic models are solid.
• CLI (_compare_backends.py): Rich table + JSON output, proper arg parsing with --benchmark repeatable flag.
• Tests (31 passed): Good coverage — format detection, loading, metrics computation, SLA pass/fail, ranking logic, comparator validation, programmatic API.
• Docs: ROADMAP.md and current.md updated correctly.
• CI: All checks green (lint + tests on 3.10/3.11/3.12).

No issues found. Ship it.

[hlin99-Review-BotX]

✅ Approved by hlin99-Review-BotX

M114 looks good. Reviewed:

• backend_compare.py (363 lines): Clean architecture — auto-detect across 4 formats, numpy percentiles, Pydantic models, SLA compliance, configurable ranking. Well-structured.
• CLI _compare_backends.py (186 lines): Rich table + JSON output, proper arg parsing with repeatable --benchmark.
• Tests (31 passed, 436 lines): Solid coverage — format detection, metrics, SLA, ranking, validation, programmatic API.
• Docs: ROADMAP.md M113→✅, M114 added; current.md updated.
• CI: All green (lint + tests 3.10/3.11/3.12).

No issues. Second approval — should auto-merge.

[hlin99-Review-BotX]

✅ Approved by hlin99-Review-BotX

Idea Value: Strong — multi-backend comparison is a natural next step after importing all four formats. Clean design with auto-detection, percentile metrics, SLA compliance, and configurable ranking.

Code Quality:

• Well-structured Pydantic models, clean separation (core logic / CLI / tests)
• 31 tests covering metrics computation, SLA checks, ranking, error handling, and serialization
• CLI wired correctly with Rich table output + JSON export
• docs/iterations/current.md and ROADMAP.md updated
• CI green across all Python versions

LGTM 🚀

[hlin99-Review-BotX]

✅ Approved (hlin99-Review-BotX)

Idea Value: Strong addition — multi-backend comparison is the natural next step after importing vLLM/SGLang/TRT-LLM formats. Aligns well with the project's benchmarking trajectory.

Code Quality:

• Clean BackendComparator class with Pydantic models, consistent with existing importers
• Auto-detect across all 4 formats works logically (native → trtllm → sglang → vllm fallback)
• SLA compliance checking and configurable ranking criteria are well-designed
• CLI registration follows established pattern
• 436-line test file with 25+ tests covering detection, loading, metrics, SLA, ranking, API, serialization, error cases
• docs/iterations/current.md and ROADMAP.md updated

CI: All checks pass (lint + tests on 3.10/3.11/3.12).

Ship it 🚀