Fail closed when Grok STT words is not a JSON array

[jmjava]

Problem

_grok_stt treated parsed.get("words") or [] as a list and skipped any
non-dict item. A string words field was iterated character-by-character, and
a scalar/list item inside words/segments silently dropped that token. Whisper
word indices and wait_word then pointed at the wrong spoken times.

This is the leftover from PR #146 (_stt_json_number on start/end did not
cover the list shape).

Fix

• Present words / segments must be JSON arrays.
• Each item must be a JSON object (words[i] / segments[i]).
• Missing or null words still becomes [] (no word-level timestamps).
• Empty segments: [] still synthesizes a segment from words (or duration).

Tests

• test_grok_stt_string_words_raises
• test_grok_stt_non_object_word_raises
• test_grok_stt_object_segments_raises
• test_grok_stt_missing_words_still_ok

Local: ruff check src/ tests/ green; pytest tests/ -v --tb=short 825 passed, 1 skipped;
docgen benchmark meets src/docgen/benchmark_data/baseline.json (no --update-baseline).

Out of scope

• Empty-token word rows (word: "") are still skipped; remaining token indices stay dense.
• OpenAI whisper-1 SDK path (verbose_json typed .words / .segments).
• Issue scene-spec / Manim beats drift from narration (label→wait_word mismatch + unpaced timed_play) #56 (label → wait_word semantic matching).

  

[cursor]

Stopping this loop. This is another sibling fail-closed type check, not product progress. Do not reopen. Do not mint #150.