Fix backend logic bugs: 13 findings

[JMRussas]

Summary

• 13 findings from the codebase evaluation, covering 3 HIGH and 10 MEDIUM/LOW severity issues
• Extract shared JSON utilities (_extract_json_object, _strip_trailing_commas) to backend/utils/json_utils.py for reuse across planner, verifier, and knowledge extractor
• Fix quadratic token growth in Claude agent by pruning conversation history to last N rounds (configurable max_history_rounds)
• Fix context forwarding race condition by wrapping per-dependent read-modify-write in DB transaction
• Fix budget leak on plan parse failure — API spend is now recorded even when JSON parsing fails
• Fix verification feedback cap: proper sliding window instead of remove-all-then-append
• Include RUNNING tasks in cancel_project so in-flight work gets cancelled
• Add retry limit check to review_task retry action (was uncapped)
• Switch to paragraph-based requirement numbering for multi-line requirements
• Skip budget check for Ollama-only projects (free tasks shouldn't pause on budget exhaustion)
• Add API timeout to verifier and knowledge extractor calls (were missing)
• Skip verification/extraction when budget is exhausted (log warning, don't block)
• Truncate long output before sending to verifier (cap at 8000 chars)
• Add _reset_warned_models() with autouse test fixture to prevent state leak
• Make HTTP client timeout configurable (execution.http_client_timeout)

Test plan

• 22 new tests in tests/unit/test_backend_logic.py covering all findings
• Updated 2 existing test files for paragraph-based requirements format
• Full suite: 701 passed, 0 failed
• Lint: all checks passed

Generated by Claude Code · Claude Opus 4.6