Add a 'How Dayflow figured this out' transparency panel

[caezium]

What this adds

A "How Dayflow figured this out" panel, shown inline in each activity card's detail (collapsed by default). It makes the AI categorization transparent — useful for trust in an always-on screen recorder, and for debugging odd cards. For each card it surfaces:

• The model's reasoning for how it grouped/categorized the session.
• The on-screen observations it wrote, each timestamped and paired with a downsampled screenshot from that moment — so you can see what the model actually saw.
• Which model/provider generated the card.
• An optional raw LLM-call view (operation, status, latency, response).

How it works

• Reasoning is persisted in the timeline_cards metadata JSON at save time — no schema migration; it rides the existing envelope alongside appSites. For older cards (pre-this-change), it falls back to parsing the logged llm_calls, so the panel is populated retroactively too.
• Observations, screenshots, and raw calls are read from already-persisted data by batchId. No new tables.
• Screenshot decoding is downsampled via ImageIO and loaded off the main thread.

Tests

Unit tests for the reasoning extraction (CardInsightParsingTests) covering bare JSON, code-fenced JSON, and the OpenAI / Gemini envelope shapes, plus graceful nils.

Notes

• Branch is based on current main (v1.14.1); the app builds and the tests pass.
• Screenshots: the per-observation screenshots are the most compelling part ("see what it saw"), but if you'd prefer to land a text-only version first for privacy reasons, that split is easy — happy to adjust.
• I'll add a screenshot of the panel to this PR.