Provide a reference scenario runner in `examples/`

[AshwinUgale]

Context

From the README: "goldset builds, versions, and checks scenarios; an external runner replays the turns against your system." No reference runner ships today, so for newcomers the loop is only closed on paper.

Task

Provide examples/run_scenarios.py that:

1. reads a scenario bundle,
2. replays each turn's input against a pluggable "system" callable, and
3. writes a transcript that goldset scenarios check accepts.

Ship a deterministic offline stub system (so CI can run it) and document how to swap in an OpenAI-compatible endpoint behind a clearly opt-in path.

Acceptance criteria

• Offline stub run is deterministic and exercised by a test.
• No network dependency in the core or default test path.
• OpenAI-compatible mode is opt-in and clearly separated (no import at module top level that requires network/keys).
• examples/README.md documents both modes.

Pointers

• src/goldset/scenarios/
• examples/

Help wanted

Needs a small design decision on the runner interface (how a "system under test" is passed in). Pairs naturally with the transcript-format docs issue.