Fail closed when timing.json words/segments are not object arrays

[jmjava]

Problem

PR #126 required each timing.json stem to be a JSON object. Inner words / segments were still untyped.

A stem like {"words": "x"} or {"words": ["hello"]} then:

1. Compile / scene-spec-generate coerced a non-list to [] (pace: none compiled as if timestamps had never been run).
2. A list of non-objects kept LLM wait_word indices and later crashed Manim on .get("start").
3. sync_audio_tail_waits_in_scenes treated a truthy non-list segments string as present and tried to patch scenes.py.

Change

When words or segments is present and not null, load_bundle_timing requires a JSON array of objects. Missing / null / [] stay allowed (empty still means no timings; paced compile already fails).

Tests

Local: ruff check src/ tests/ green; pytest tests/ 734 passed, 1 skipped; docgen benchmark meets src/docgen/benchmark_data/baseline.json (no --update-baseline).

• Reject string / object / int words and segments
• Reject words: ["hello"] and segments: [1]
• Accept missing / null / empty arrays
• extract_all does not rewrite the file
• pace: none compile, validate timing_sync, wizard statuses, scene-asset checks, and sync_audio_tail_waits surface TimestampError

Out of scope

• Empty segments.all still leaves timing.json unchanged (Fail-closed timestamps, concat, and Anthropic base_url #78)
• Requiring numeric start / end on every word row
• Bootstrap _load_timing helpers inside compiled scenes.py