v3.6.0

Medium-tier audit fixes (cross-platform, correctness, stability, security).

Fixed

• OpenAI o-series (o3/o3-mini/o4-mini) now send max_completion_tokens instead of max_tokens (the o-series 400 on the latter).
• OLLAMA_BIN was macOS-only — resolved via $OLLAMA_BIN → PATH → app bundle, so /pull, /delete, and update checks work on Linux/Windows.
• multi_edit failed in Ollama text mode — JSON-string edits are now parsed; no longer burns an undo slot on a no-op.
• _stop_ollama killed all Ollama processes — now stops only the one Kodiqa started and reaps it; added an atexit handler to clean up spawned MCP/LSP/Ollama children even on crash.
• memory_search could fail under parallel dispatch — SQLite opened with check_same_thread=False.
• Streaming connection leak — all three chat loops resp.close() in finally.

Security

• Hook command injection — hook {param} values are shlex.quoted before substitution.
• Blocklist hardening — run_command normalizes whitespace before matching BLOCKED_COMMANDS (rm -rf / is now caught).

Tests

• Regressions for o-series body, multi_edit JSON-string, and blocklist normalization (300 total).