Refactor tool infrastructure, per-request LLM configuration, and add LLM resilience

[Hemdan47]

This PR introduces significant architectural improvements to the agent's tool infrastructure, LLM configuration, and error-handling resilience, enhancing flexibility, robustness, correctness, and maintainability.

Changes

Migrate to per-request LLM configuration via LiteLLM

• Removed global config.py LLM configuration.
• Introduced llm/factory.py with dynamic model instantiation per request.
• Added llm/helpers.py with utility functions for LLM operations.
• Updated API schemas to support per-request model configuration.
• Modified health check endpoints to reflect new architecture.

---

Features

Add centralized Tenacity-based retry policy for LLM resilience

• Introduced a custom _ResilientRunnable wrapper using Tenacity @retry to handle LLM execution errors.
• Centralized shared Layer-2 tenacity configuration in retry_policy(), which handles retryable exceptions, exponential backoff with jitter, and emitting transient llm.error events.
• Applied the centralized retry_policy to wrap the ReAct loop in the analyzer subgraph (run_analyzer), ensuring standard retries and notifications cover the entire agent invocation.

Replace hardcoded skip-lists with .gitignore-based filtering

• Added new ignore_filter.py module using pathspec library for .gitignore parsing.
• Glob, grep, and list tools now respect .gitignore patterns automatically.
• Removes maintenance burden of hardcoded directory/file exclusions.
• Always skips .git directories regardless of .gitignore settings.
• Comprehensive test coverage for ignore filtering logic.

Configurable working directory for file tools

• Converted glob, grep, and list tools to BaseTool classes.
• Added working_directory parameter for flexible path scoping.
• Improved tool consistency and reusability across different contexts.

LLM resilience and targeted exception handling

• Implemented targeted LLM exception handling to make failures self-heal and visible instead of crashing opaquely.
• Introduced classification of LLM failures into actionable categories (e.g., rate limits, bad requests, authentication errors, timeouts, service unavailable).
• Introduced a new "llm.error" event to broadcast transient and fatal LLM errors to downstream clients.
• Guarded I/O boundaries against system/disk issues (e.g., writer disk failures, glob/grep errors).
• Added warning filter in pyproject.toml to ignore cosmetic Pydantic serialization warnings.

---

Bug Fixes

Prevent duplicate title emissions in planner

• Fixed state management issues causing repeated title broadcasts.
• Ensures cleaner streaming output during planning phase.

Classify mutating git subcommands correctly

• Bash tool now properly categorizes git commands like add, commit, push as modify operations.
• Improves permission request accuracy for git operations.

Graph Recursion Error Handling

• Caught GraphRecursionError in the runner and emitted a clear "stuck in loop" event message (run.failed with step budget warning) instead of a generic crash.

---

Testing

• Added 15 new unit test cases for the LLM factory (test_factory.py).
• Added 7 new unit tests for the ignore filter functionality (test_ignore_filter.py).
• Expanded unit tests for the glob, grep, and list tools to validate working_directory scoping, .gitignore filtering, and edge-case error handling (e.g., dangling symlinks and malformed regex).
• Added unit tests for tenacity retry execution paths, event hooks, and fatal error propagation (test_resilience.py).
• Introduced tests/integration/helpers.py to unify LLM config resolution and provider key extraction.
• Refactored planner, generator, analyzer, and stream router integration tests to use the new unified helper, eliminating duplicated boilerplate.
• Updated integration tests for streaming routes, planner, and analyzer.

---

Dependencies

• Added pathspec dependency for .gitignore pattern matching (pyproject.toml).