Harden LLM deserialization and fallback handling

[lethanhson9901]

Summary

This PR improves resilience when model outputs do not strictly match expected schemas during documentation generation.

It addresses failures like:

• Failed to deserialize ... missing field 'name'
• Failed to deserialize ... missing field 'responsibilities'
• Similar missing/loose field issues in AI analysis outputs

Root Cause

Some LLM responses are schema-unstable (missing required fields, variant naming drift, mixed value shapes such as object/string/array), while parts of the pipeline previously required strict deserialization.

Changes

• Hardened deserialization for code insight structures with lenient parsing and defaults.
• Added tolerant mapping for CodePurpose from raw/variant text.
• Added robust parsing for AI code-purpose analysis fields (code_purpose, confidence, reasoning).
• Added fallback behavior in code insight analysis:
  • if AI extraction fails, return static-analysis insight instead of failing the whole run.
• Added #[serde(default)]/Default handling for key report structures to avoid hard failures on partial outputs.
• Added Mermaid safety constraints to relevant compose prompts to reduce invalid Mermaid syntax generation.

Tests

• cargo test passed (53 passed, 0 failed).
• cargo check passed.
• Verification repeated in a clean temporary worktree with only this patch applied.

Scope / Out of Scope

• Scope is intentionally limited to deserialization robustness, fallback safety, and Mermaid prompt hardening.
• Large unrelated local workspace changes were intentionally excluded to keep this PR safe and reviewable.

Impact

• Prevents pipeline interruption from non-critical LLM schema variance.
• Preserves documentation generation continuity by degrading gracefully to static analysis when needed.
• Improves output reliability without changing core workflow intent.