fix: Remove remaining MLflow references after migration to OpenTelemetry (#1422)

Completes the cleanup from PR #1420 by removing all lingering MLflow references:

- Remove MLflow tracking setup section from setup.sh (lines 605-620)
- Update README.md to reference OpenTelemetry observability stack
- Update snowball-method.md to reference OpenTelemetry instead of MLflow
- Delete mlruns/ artifact directory from repository

The observability/README.md intentionally retains MLflow references as
historical migration documentation.

Closes #1421

Author: atxtechbro

README.md

@@ -69,7 +69,7 @@ This system enables **macro-level agent management** instead of micro-level file
 
 - **Harness-Agnostic Configuration**: Single `.agent-config.yml` defines user preferences, agent settings, and paths - works across Claude Code, Amazon Q, and Codex without duplication (see [config-architecture.md](docs/config-architecture.md))
 - **Reproducible Agent Procedures**: Slash commands in `commands/` directory (`/close-issue`, `/create-issue`, `/extract-best-frame`, `/retro`) enforce consistent workflows across all AI harnesses
-- **Telemetry and Feedback**: `bin/claude-with-tracking` wraps agent sessions with MLflow tracking for performance analysis and continuous improvement
+- **Telemetry and Feedback**: OpenTelemetry observability stack with Grafana dashboards provides real-time performance monitoring and continuous improvement insights (see [observability/README.md](observability/README.md))
 - **Parallel Execution**: Using [tmux + git worktrees](knowledge/procedures/tmux-git-worktrees-claude-code.md), you manage multiple AI agents simultaneously across parallel tasks
 - **Principle Enforcement**: `knowledge/procedures/` and `knowledge/principles/` automatically loaded into agent context to maintain consistency
 

knowledge/principles/snowball-method.md

@@ -8,6 +8,6 @@ The "snowball method" focuses on continuous knowledge accumulation and compoundi
 - Compounding returns: Small improvements accumulate and multiply rather than remaining isolated
 - Reduced cognitive load: Less need to "re-learn" or "re-discover" previous solutions
 
-With MLflow tracking, we transform ephemeral Claude sessions into queryable history. This enables kaizen continuous improvement. Each tracked session provides insights to debug failures and improve the next. We turn isolated interactions into a compounding knowledge base. Every session makes future sessions better.
+With OpenTelemetry observability, we transform ephemeral AI agent sessions into real-time metrics and queryable history. This enables kaizen continuous improvement. Each session provides insights through Grafana dashboards to debug failures and improve the next. We turn isolated interactions into a compounding knowledge base. Every session makes future sessions better.
 
 This principle ensures that our development environment continuously improves over time.

setup.sh

@@ -602,23 +602,6 @@ if [[ -f ~/.bash_aliases ]]; then
   echo -e "${GREEN}✓ Bash aliases loaded successfully${NC}"
 fi
 
-# MLflow tracking setup (self-healing with spilled coffee principle)
-echo -e "${DIVIDER}"
-echo "Setting up MLflow tracking..."
-
-# Check if start-mlflow script exists
-if [[ -x "$DOT_DEN/bin/start-mlflow" ]]; then
-  # Use the start-mlflow script which handles all checks
-  "$DOT_DEN/bin/start-mlflow" start
-  # The script handles:
-  # - Installing MLflow via uv if not present (spilled coffee principle)
-  # - Checking if already running (idempotent)
-  # - Starting in background if needed
-  # - Silent operation to avoid noise
-else
-  echo -e "${YELLOW}start-mlflow script not found. Skipping MLflow setup.${NC}"
-fi
-
 # Configure git hooks
 if [[ -d "$DOT_DEN/.githooks" ]]; then
   echo "Configuring git hooks..."