feat(engine): add export_traffic to dump observations as savings-report JSONL

[AZERDSQ131]

Problem

The savings report (rulechef-savings) needs a traffic JSONL file, but there's no helper to produce one. Observation mode (start_observing / add_observation) already captures exactly this data, it just never lands on disk in the traffic format — users had to write their own dump loop.

Fix

Added RuleChef.export_traffic(path). It exports LLM-sourced observations from the buffer as:

• {"text": ..., "llm_label": ...} for classification output
• {"text": ..., "llm_entities": [...]} for NER/extraction output (entities or spans key)

Human-sourced examples (add_example, corrections) are skipped — they aren't traffic. Observations missing a text input or a recognized output shape are also skipped rather than raising.

Documented the observe -> export_traffic -> rulechef-savings loop in benchmarks/INSPECTING_RULES.md, including a note that rulechef-savings itself currently only scores the classification (llm_label) shape (NER traffic export is ready for when #19 lands).

Tests

• Classification and NER export shapes.
• Human examples are skipped.
• Observations without text / unrecognized output are skipped (count stays 0, file stays empty).
• Full round-trip: export_traffic -> rulechef-savings CLI -> assert the generated HTML report (mirrors tests/test_reports.py::test_savings_cli's invocation pattern).

pytest tests/ -q → 316 passed, 1 skipped. ruff check and mypy rulechef/engine.py clean.

Closes #20

[adaamko]

Thanks for picking this up — unfortunately it crossed with #21, which was opened for the same issue (#20) a few days earlier and just merged. Same design, so closing this one. The spans output fallback in yours is a nice touch #21 doesn't have; if you hit a real case that produces spans instead of entities, a small follow-up PR for that is very welcome. Also relevant to your #28: main now has export_traffic, so the NER savings path can be tested end-to-end.