fix: shared LLM output guardrail layer (closes #17)

[PRATHAMESH75]

Summary

Closes #17. Generalizes the one-off <think>/code-fence fix from commit 985f789 into a single, reusable output guardrail / validation layer that every agent response flows through.

Reasoning models (MiniMax/GLM/DeepSeek/Qwen) and JSON-mode endpoints leak <think> blocks, markdown fences, and truncated/malformed JSON into the content field. Previously only LLMClient handled some of this inline, while three other paths parsed LLM JSON independently with duplicated repair logic — and the two simulation generators stripped nothing, so the same output could still cause 500s.

Changes

• New app/utils/llm_sanitizer.py — pure, testable helpers:
  • strip_reasoning (handles well-formed, truncated, and orphan-closer tag variants)
  • strip_code_fences (extracts JSON from ```json fences, incl. unclosed)
  • extract_json_block (first balanced {…}/[…] out of surrounding prose)
  • close_truncated_json (string/stack-aware closing of token-limit truncation)
  • repair_json (progressive repair passes)
  • parse_json — the single validation step: sanitize → parse → repair → optional required_keys schema check; raises one predictable ValueError instead of leaking JSONDecodeError
• llm_client.py — chat() → sanitize_content; chat_json() → parse_json (now repairs instead of raising; added optional required_keys). Covers report_agent, zep_tools, ontology_generator for free.
• oasis_profile_generator.py / simulation_config_generator.py — these bypass LLMClient; now sanitize content before parsing and delegate their duplicated _fix_truncated_json/_try_fix_*_json bodies to the shared module (oasis keeps its bio/persona salvage fallback).

Acceptance criteria

• All agent outputs pass through a single validation/sanitization step
• Malformed reasoning-model output no longer causes 500s

Testing

• New tests/test_llm_sanitizer.py: 24 tests (reasoning/fence stripping, truncated/nested JSON repair, required-keys validation, and an LLMClient.chat_json integration test proving reasoning-block + fence + truncated-brace output parses instead of 500ing).
• Full backend suite: 28 passed. The single failure (test_sentiment_analyzer_cache) is pre-existing and unrelated (a _classify_batch tuple-arity bug in sentiment_analyzer.py); it fails identically on a clean tree.

🤖 Generated with Claude Code