2.91.0

Added

• Seed-sponsors toggles: four global settings control which LLM prompts receive the known-sponsor list (pass 1 detection, pass 2 verification, reviewer review, reviewer resurrect). All default on. Turning the reviewer toggle off makes that pass an independent second opinion instead of one primed with the same sponsor hints as the detector. Exposed in the settings API and a new Seed sponsors section in the settings UI (#683).
• Cross-episode text recurrence hints (off by default): spans of the transcript whose wording repeats near-verbatim across the feed's recent episodes are fed to pass-1 detection as a hint for intros, credits, and other boilerplate. Text evidence is a hint only and never cuts audio on its own. Complements the audio fingerprinter, which misses re-recorded boilerplate that is verbatim in words but acoustically different each week (#683).
• Experimental segment-ID addressing mode (timestamps remains the default): the detector can ask the model for numbered transcript line IDs instead of absolute timestamps, then map the IDs back to exact Whisper segment times. An invented timestamp is undetectable; an invented ID is provably out of range, so bad references are dropped instead of cutting the wrong audio. Falls back to timestamp parsing when a model ignores the ID contract. Selectable under Experiments in the settings UI; benchmark results decide whether the default ever changes (#683).

Documentation

• New glossary entries and configuration sections for seed sponsors, text recurrence hints, and the addressing mode; openapi.yaml covers the six new settings keys (#683).