feat: add opt-in AgentPond tracing

[marcusschiesser]

What changed

• adds opt-in OpenInference LLM spans around PentesterFlow's real OpenAI-compatible chat and streaming path
• exports spans through AgentPond's environment-driven Files SDK exporter
• records structural request/response metadata only; prompts, responses, targets, and credentials are excluded
• documents local setup and keeps .agentpond/ state out of git

AgentPond fits PentesterFlow's audit-oriented workflow because operators can keep trace storage local or select their own Files SDK provider. Nothing is exported unless AGENTPOND_ENABLED=true and a Files SDK environment is configured.

Validation

Commands run:

npm ci --dry-run
npm run typecheck
npm run lint
npm run test
npm run build

All passed: 62 test files / 657 tests, plus the production bundle.

End-to-end trace validation used the existing in-process OpenAI-compatible server fixture, which exercises the actual OpenAIClient HTTP and SSE paths:

AGENTPOND_ENABLED=true \
FILES_SDK_PROVIDER=fs \
FILES_SDK_ROOT="$PWD/.agentpond/envs/local/objects" \
AGENTPOND_PROJECT_ID=default-project \
npm run test -- src/llm/openai.test.ts

npx agentpond sync
npx agentpond traces list --limit 10 --json
npx agentpond traces get 794674826ab0764c2a8c0d6edfbbc8cb --json

The focused integration suite passed all 17 tests. AgentPond processed 15 objects / 30 events, and trace 794674826ab0764c2a8c0d6edfbbc8cb read back as openai-compat.chat with model qwen-coder, OpenInference span kind LLM, and privacy-safe request/output counts.

Limitations

This first integration covers OpenAI-compatible backends (including LM Studio, Kimi, OpenRouter, and DeepSeek routing through that client). The separate Ollama, Gemini, and Anthropic clients are unchanged.