feat: use OpenRouter directly from commit script

[wajeht]

Summary

• use one OpenRouter API key and OpenAI-compatible inference endpoint
• ask first-time users only for a hidden OpenRouter API key
• create the config directory and file with 700 and 600 permissions
• support rerunning configuration with --setup
• support API-key and model overrides through environment variables
• send Git diffs directly from the Bash client to OpenRouter
• keep authorization data out of process arguments and clean up private temporary files
• fail public download commands on HTTP errors and bound inference requests with timeouts
• remove the server-side AI proxy, IP allowlist, and server-owned provider keys
• remove separate Gemini/OpenAI configuration and Make targets
• add a comprehensive manual security and release QA checklist
• ask for confirmation by default, with explicit --yes automation support
• update the usage guide and integration coverage

Why

Users should need only one inference account and API key. commit.jaw.dev should only distribute the script instead of receiving credentials or repository diffs.

User impact

Run curl -fsSL https://commit.jaw.dev | bash. The first run asks only for an OpenRouter API key, then creates ~/.config/commit/config.json with private permissions. The script automatically uses google/gemini-2.5-flash-lite; advanced users can override it with --model or COMMIT_MODEL. Use --setup to replace the saved key.

Use --dry-run to preview without committing. Real commits ask for confirmation by default; use -y or --yes to explicitly accept without prompting. Commit always passes --no-verify to Git, so hooks are skipped without requiring a public flag.

Validation

• bash -n assets/sh/commit.sh
• go test ./...
• go test -race ./...
• go vet ./...
• git diff --check