Fix AttributeError crashes when LLM returns malformed JSON

[martindecz]

Summary

Fixes two AttributeError crashes that occur when the LLM returns malformed JSON output, causing extract_json() to fail and downstream functions to receive wrong types.

Closes #199

Changes

1. pageindex/utils.py - extract_json() now returns [] (empty list) instead of {} (empty dict) on parse failure. All callers expect a list of dicts, so the fallback value should match that contract.

2. pageindex/page_index.py - process_no_toc() now validates that generate_toc_init() and generate_toc_continue() return lists before calling .extend(). If they return a non-list type (e.g. dict from failed JSON parsing), the code logs a warning and either resets to an empty list or skips the result.

3. pageindex/page_index.py - meta_processor() list comprehension now filters out non-dict items with isinstance(item, dict) before calling .get(), preventing 'str' object has no attribute 'get' errors.

Reproduction context

These crashes are reliably triggered when using a local vLLM endpoint with smaller models that frequently produce malformed JSON (extra data, empty responses, malformed structure). All 7 test documents (Office docs converted to PDF) failed with these errors.

Test plan

• Process PDFs with a model that produces clean JSON output (no regression)
• Process PDFs with a model that produces malformed JSON output (crashes prevented, graceful degradation)

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[martindecz]

Closing this PR — our type guards were based on incorrect assumptions about the logger interface (JsonLogger doesn't have .warning()) and the fallback return type change from {} to [] broke more call sites than it fixed (6 out of 13 callers expect dict). The root cause is better addressed in extract_json() itself with robust JSON repair, not in individual callers. Apologies for the noise.