fix: use unique temp file in SKILL.md phase2 instructions to avoid parallel overwrites

feat: Add logical fallacy detection with 12 fallacy types, CLI flag (--fallacies), and cost estimation support.

feat: Add real-time SSE streaming for Phase 1 with parallel and sequential modes.

feat: Add quorum voting (--vote), multi-round dialogue (--rounds N), and bias tracking (bias command).

feat: Implement fallback model support for providers, including configuration and status reporting.

feat: Add two-pass Phase 2 execution for deep mode, enabling deferred cross-critique for local models via a new `phase2` CLI command.

feat: Add x.ai provider support by implementing `_call_xai` and integrating it into the model calling mechanism with corresponding tests.

feat: Add EMA-based model scoring and leaderboard functionality with a new CLI command.

feat: Implement session context building with token budget management and gracefully handle null content responses from LLM providers.

refactor: Use a shared httpx.AsyncClient for all API calls to improve connection reuse.

feat: Implement JSON-based ranking extraction with re-prompting and regex fallback for critique parsing.

feat: Add initial test suite for orchestrator, providers, and ranking modules, and update README.

docs: update README to reflect new modular package architecture and features

feat: Modularize Conclave application by introducing dedicated modules for sessions, configuration, progress reporting, and more.

docs: update README with new sections for cost estimation, multi-turn sessions, real-time progress, and retry logic.