TL;DL v2.5.0 — Editorial decks everywhere and a new transcription model

The first TL;DL release in four months. Episode summaries now carry their editorial deck and pull quote into the email newsletter and the RSS feed, transcription runs on OpenAI's new gpt-transcribe model, and a season of pipeline reliability fixes ships alongside.

What's new

• Summary emails now open with the episode's editorial deck and feature its pull quote, matching the broadsheet look of the site.
• The RSS feed leads each item with the episode's deck, so feed readers see the standfirst instead of a raw summary excerpt.
• Transcription moved to OpenAI's gpt-transcribe model — their recommended successor to the 4o-transcribe family. Transcripts now reliably run to the end of the episode (the old model could stop early on some files), and the whisper-1 safety net for unusual audio is unchanged.
• The admin dashboard was redesigned into a tabbed layout, including a new subscribers view.
• Shared links look better: a proper 1200×630 social card, and page titles that lead with the tagline.

Fixes

• Curly quotes and other special characters in episode titles no longer show as raw “-style codes — new titles are decoded at ingest, and existing ones were cleaned up.
• An episode whose final minutes can't be decoded now keeps its transcript (marked partial) instead of failing outright.
• Long episodes hosted on rate-limiting CDNs (looking at you, Substack) now ride out the throttling instead of failing mid-transcription.
• Episodes no longer show up twice when the monitor and a manual submission both pick them up.
• Adding a podcast whose directory listing is incomplete can no longer queue up its entire back catalog by mistake.

Under the hood

• Queue processing hardened: zombie messages are dropped instead of retried forever, and the consumer's CPU budget was raised for very long episodes.
• Test stack upgraded to vitest 4 + Workers pool 0.15, CI now runs on every push, and Hono security patches applied.
• A summary-model A/B harness compares candidate models on real transcripts before any model swap.
• The Postmark email templates are now versioned in the repo as source-of-truth copies.