[Feature] Unit tests for jsonblock JSON extraction and truncation recovery

[sjohnston1972]

Part of the "Add an automated backend test suite" parent issue. Depends on the pytest-scaffolding sub-issue.

Background

backend/app/jsonblock.py extracts a JSON object from raw LLM output. It has four strategies (fenced ```json block, any fenced block, last balanced {...}, and a brace-counting truncation-recovery pass in `_try_truncation_recovery`). All designer/critic/orchestrator/extractor parsing depends on it, so a regression here silently breaks every agent.

Problem / Goal

The module is pure (no I/O) and completely untested. Add thorough unit tests.

Where to look

• backend/app/jsonblock.py — extract_json (entry point) and helpers _clean, _try_parse, _try_truncation_recovery, _last_balanced_object.
• The module docstring lists the exact cases it's meant to handle (nested objects, trailing commas, // and /* */ comments, truncation). Use it as the test checklist.

Suggested approach

Write backend/tests/test_jsonblock.py covering at least:

1. Clean fenced ```json block → parsed dict.
2. Unfenced JSON with trailing prose after it → last balanced object wins.
3. Nested objects like {"citations": [{"page": 42}]} → returns the full object, not the first inner } (the bug the module was written to fix).
4. Trailing commas before }/] and // / /* */ comments → recovered via _clean.
5. A brace inside a string value, e.g. {"note": "value with } brace"} → not fooled by the }.
6. Truncation: input like {"a": 1, "b": [{"c": 1}, {"c": 2}, {"c": 3 → recovers {"a":1,"b":[{"c":1},{"c":2}]} (the docstring's own example).
7. Empty string / no JSON at all → returns None.

Acceptance criteria

• test_jsonblock.py covers all seven cases above and passes.
• Each of the four extraction strategies is exercised by at least one test.

Testing

cd backend && pytest tests/test_jsonblock.py -q.

Out of scope

• Changing jsonblock.py behavior — this is characterization/coverage only. If a test surfaces a genuine bug, file it separately.