ElevenLabs eleven_v4: no speed setting, timestamps work, voices need pacing fixes

eleven_v4 supports /v1/text-to-speech/{id}/with-timestamps (alignment includes the [audio tag] characters at tiny durations, so strip bracketed spans before building word timings). v4 has only stability and similarity settings: no speed or style, and no SSML breaks, so delivery runs slow (~2.3 words/s) for short-form video. What worked: ffmpeg atempo 1.10 (1.15 on the opening line), shortening internal silences over 0.22 s to 0.16 s, scaling alignment times by 1/tempo plus a cut map, and 0.12-0.2 s gaps between lines. Checking the result with Scribe word timestamps gave a median word-start error of about 35 ms. Library PVC voices list eleven_v4_post_trained in high_quality_base_model_ids, and the v4 docs warn that Voice Design voices are less performative, so prefer post-trained library voices. v4 billed about 0.45 credits per character.