stock-watchlist-agent: add inline experiments (capture/replay regression testing)

[mehulsonowal]

Layers the ddtrace inline experiments workflow onto the stock-watchlist-agent — "a unit test for LLM apps": mark an input→output boundary, capture a baseline of current behavior, edit code, replay the current code against it, and get per-case match/changed verdicts (plus optional evaluators and a --publish path to the Experiments UI).

What it adds:

• @experiment_start on the analyze_portfolio boundary + attached evaluators (completeness + sentiment/grounding LLM judges)
• capture_watchlists.py / capture_quotes.py baseline harnesses
• INLINE_EXPERIMENTS_GUIDE.md — a concise capture → replay → --evaluate → --publish quickstart
• .env auto-loading (sitecustomize.py) + an observability init helper

Loop:

cd test-apps/stock-watchlist-agent && python -m venv .venv && .venv/bin/pip install -e .
export OPENAI_API_KEY=sk-...
exp() { PYTHONPATH=. .venv/bin/ddtrace-experiment "$@"; }
exp capture capture_watchlists:generate_traffic
exp replay  src.agents.orchestrator --evaluate

Requires a ddtrace version that includes inline experiments (the ddtrace-experiment CLI and ddtrace.llmobs.experimental). Currently pre-release — do not merge ahead of that ddtrace release.

🤖 Generated with Claude Code