Raw Bedrock model IDs, Opus 4.7/4.8, /swe + /summarize skills, uv-managed deps

[aarora79]

Summary

This PR makes two complementary changes that map to the two purposes the repo now serves: run non-Anthropic models through Claude Code, and measure how well each model actually does coding work.

1. /swe and /summarize skills + worked example

• Adds .claude/skills/swe/SKILL.md — drives any Claude Code model through a real software-engineering task in any GitHub repo, producing four artifacts (github-issue.md, lld.md, review.md, testing.md) under benchmarks/swe-benchmark-data/{repo}/{problem}/{model}/. The skill stops at design — it does not implement the change.
• Adds .claude/skills/summarize/SKILL.md — post-run report covering artifact completeness, error signals, token usage broken down by model + cache type, and recurring themes.
• Ships a worked example targeting agentic-community/mcp-gateway-registry at tag 1.24.4 with two scoped problems (remove-faiss, remove-efs-from-terraform-aws-ecs). qwen.qwen3-coder-next already has the four artifacts on remove-faiss so reviewers can see expected output before running anything.
• The cloned target lives at benchmarks/swe-benchmark-data/{repo}/repo/ and is gitignored — each contributor pins their own checkout rather than this repo carrying large third-party trees.

2. Raw Bedrock model IDs end-to-end + Opus 4.7/4.8 + uv-managed deps

• Drops the alias layer across claude-model.sh, litellm-config.yaml, humaneval_runner.py, and bedrock/README.md. What you pass to --model is now the same raw Bedrock model ID that hits the wire (e.g. qwen.qwen3-coder-next, us.anthropic.claude-opus-4-8).
• Adds Claude Opus 4.7 and 4.8 to the native Bedrock list (verified ACTIVE via aws bedrock list-inference-profiles). Counts go 43 → 45 total, 5 → 7 native.
• Audited the OpenAI-compat Bedrock endpoint (bedrock-mantle.us-east-1.api.aws/v1/models) — the 38 third-party entries match the catalog exactly, nothing missing.
• Adds bedrock/pyproject.toml so uv sync works directly. Replaces the runtime pip install calls in setup-proxy.sh that broke under PEP 668 on Python 3.12+. Pins requires-python = ">=3.10,<3.14" because uvloop 0.21 cannot import on Python 3.14.
• setup-proxy.sh now binds to 127.0.0.1 by default with an optional --host override (prints a warning when non-loopback). The proxy does not authenticate clients itself, so the previous 0.0.0.0 default was unnecessarily exposed.
• Date-stamps "latest"/"newest" descriptions in the picker so recency claims are falsifiable; relabels MiniMax M2.5's 80.2% SWE-bench number as vendor-claimed.

3. README reframe (root + bedrock)

• Root README.md reframed around the two purposes. New Evaluation 1 — SWE skill and Evaluation 2 — HumanEval sections, with a clear note that "SWE" here means software engineering generally and is not SWE-bench.
• Repository structure block updated for .claude/ and benchmarks/.
• bedrock/README.md path fixes: LICENSE, clone URL, "See Also" link, alias paths — all previously pointed at the old shekharprateek/... standalone repos that this monorepo replaced.

Test plan

• cd bedrock && uv sync succeeds on Python 3.10–3.13
• ./scripts/setup-proxy.sh starts cleanly, binds to 127.0.0.1:4000, curl http://127.0.0.1:4000/health returns 200
• ./scripts/setup-proxy.sh --host 0.0.0.0 prints the network-exposure warning and binds to all interfaces
• ./scripts/claude-model.sh --list shows 45 models with the date-stamped "latest" tags
• ./scripts/claude-model.sh --model us.anthropic.claude-opus-4-8 -p "hi" succeeds (after enabling Opus 4.8 in Bedrock model access)
• ./scripts/claude-model.sh --model qwen.qwen3-coder-next -p "hi" succeeds (with proxy running)
• All local markdown links in root + bedrock READMEs resolve
• /swe skill end-to-end on a fresh problem produces all four artifacts under the expected path