feat: 0.2.0 add drop-in SDK adapters, examples, benchmarks, fuller tests

[EvanPaules]

Adds an adoption-focused release without changing the core ContextOptimizer API. Wraps the OpenAI and Anthropic SDKs so a one-line swap turns any existing chat call into an auto-trimmed one, and ships the runnable examples, benchmarks, and tests reviewers expect to see before they star a library.

• src/adapters/openai.ts + src/adapters/anthropic.ts: withOptimizer(client, config) returns a client whose chat.completions.create / messages.create intercepts messages and runs ctx-opt before forwarding. The Anthropic adapter hoists system for the optimizer view and splits it back out for the API. Both expose client, optimizer, lastMeta.
• Subpath exports ctx-opt/openai and ctx-opt/anthropic via tsup multi-entry build. openai and @anthropic-ai/sdk are optional peer deps.
• examples/: openai, anthropic, summarizer-with-real-Anthropic, langchain.
• benchmarks/run.ts: reproducible head-to-head of all 4 strategies. Results pasted into README.
• tests/strategies/relevance.test.ts + hybrid.test.ts: fill the coverage gap. tests/adapters/*: full adapter suites with mock SDKs. 21 to 52 tests.
• CHANGELOG.md following Keep-a-Changelog. README updated with the adapter quick-start, benchmark table, and changelog link.
• Documents in CHANGELOG a known limitation: the hybrid pipeline shares recentWindow between phases, so the post-relevance set is rarely compressible. Summarizer falls through to its sliding-window fallback. Splitting the windows is queued for a future release.