complex-md 0.4.0

Intent: close the resolution gaps a text-level analyzer can close, work
with whatever model a user already pays for, wire into the harnesses people
actually run, verify the claims with fixtures, and write the limits down
where a reader cannot miss them.
Change: dependency resolution gains npm/pnpm workspace packages through
exports maps, per-package tsconfig/jsconfig paths with single-level
extends, package.json imports (# specifiers), vite/webpack
resolve.alias, Python package roots derived from __init__.py chains, and
correct require_relative scoping; 14 declarative resolver fixtures verify
edges and must-not-edges per ecosystem. Model calls go through one provider
table (anthropic, openai, openrouter, gemini, xai, deepseek, mistral, groq,
together, ollama, custom OpenAI-compatible base URL); the first configured
key wins, --provider and --model provider/id override, and providers
without a pinned default require an explicit model rather than a guess.
Wiring is a per-target table: windsurf, cline and roo rule files join
claude, cursor, openhands and codex; openclaw and hermes are global
registries, opt-in via --for, configured through their own CLIs with
mcp --root <repo>. The Codex MCP entry uses the computed invocation instead
of a hardcoded npx. The spec gains "The boundary of the analysis". Tests 9 to
36. The site mirrors agents.md (MIT, credited): palette, typography, dark
mode and section order, with 4px radii throughout; author attribution in
metadata, JSON-LD and the footer on every page.