Trustworthy context capture and a faster sidepanel

This release fixes a crash loop that made the sidepanel unresponsive, ensures hyperlinks are preserved when you capture page context, and makes captured context much easier to understand and trust. Page, screen, and selection context now behaves like file attachments: it sits in the same staging area before you send, clears when the message goes, and shows up in your message history. Any context or attachment chip is clickable to inspect exactly what text was sent to the model — useful when page extraction produces noise or you want to verify a capture before a sensitive query. The sidepanel also opens noticeably faster: the AI model list is now served from a local cache on every reopen, so the panel is ready to use without waiting on a network round-trip.

What's new

• Styling updates: Updated the font to use the same as the main website, improved tab management
• Sidepanel no longer blocks on a network fetch when opening. The AI model catalog is persisted to local storage after the first load and served from there on every subsequent open.
• Updated the model catalog with Gemini 3.5 Flash, Gemini 3.1 Flash-Lite, and GPT-5.4 nano. Deprecated the preview aliases those stable models replace, and fixed the OpenAI default model from the deprecated GPT-4o mini to GPT-5.4 mini.
• Fixed an infinite injection loop that could make the sidepanel unresponsive on certain pages.
• @selection captures now preserve links, headings, bold/italic, code, and list structure as markdown.
• @page and @screen captures now emit links as markdown instead of dropping them.
• JIRA ticket descriptions and Slack message content are converted from HTML to markdown so links and formatting are preserved.
• Context chips are shown in the same pre-send staging area as file and image attachments, and are consumed per message rather than persisting across turns.
• Clicking any chip opens a content inspector showing the full text, source URL, character count, capture time, and a Copy button.