Fail closed when whisper prompt caps coerce bools to 1 - #144

Problem

Config.from_yaml already requires YAML numbers for whisper prompt caps (#117). Timing enrichment type-checks max_whisper_segment_text_chars (#122). The two count caps still did:

max_seg = int(root.get("max_whisper_segments_in_prompt", 0) or 0)
max_words = int(root.get("max_whisper_words_in_prompt", 0) or 0)

bool is a subclass of int: max_whisper_words_in_prompt: true became 1 and truncated the word stream the scene-spec LLM uses for wait_word. Quoted "12" became 12 instead of raising.

Called out as out of scope of #142.

Change

• Shared _nonneg_yaml_int for segments / words / chars caps
• Present bool or non-number raises SceneGenerationError
• Missing/null still defaults (0 = full stream; chars 200)
• Explicit 0 still lists every word

Gate

• ruff check src/ tests/ green
• pytest tests/ — 804 passed, 1 skipped
• docgen benchmark meets src/docgen/benchmark_data/baseline.json (no clock change; no --update-baseline)

Out of scope

• Config.from_yaml already gates these keys for CLI load
• Validate _check_streams / _check_drift ffprobe returncode
• Whisper API start/end float(... or 0.0) in ai_client

Milestone: milestones/whisper-prompt-caps.md