v0.2.0 Release Notes

What's New

Structured outputs: LLM responses are now constrained to a strict JSON schema ({specs: [{key, value, children}]}) via Anthropic tool-calling and OpenAI json_schema mode — no more parsing free-form JSON, so results are more reliable and nested spec groups are supported.
HTML → Markdown input: HTML strings, blobs, and fetched pages are auto-detected and converted to Markdown (via Turndown + GFM) before extraction, so table-heavy product pages extract cleanly.
Flattening options: new flattened + inheritance extract options to flatten nested specs — either keep parent-key prefixes (·) or collapse to innermost leaf pairs.

Legacy flat JSON {key: value} responses remain supported alongside the new envelope format (validator accepts both).
Prompt tightened: values always strings with units, labels preserved, empty result returns {"specs":[]}.

New test fixtures: real Axis network-camera PDFs (German) plus HTML, provider, validator, and extractor coverage.
README documents HTML input handling.