[Observability 2/10] Add metrics to MCP tool handlers

[concaf]

Context

Instrument the 3 MCP tool handlers with invocation count, duration, and error metrics. This enables use cases B1 (error rate alerting) and C1 (latency monitoring).

Scope

Declare shared metric instruments

In internal/mcp/tools/ (e.g., in a new metrics.go or in get_processed_document.go alongside the existing shared errOAuthTokenNotFound constant):

unstructured_data_mcp_tool_invocations_total  — Counter  — attributes: tool_name, status
unstructured_data_mcp_tool_duration_seconds   — Histogram — attributes: tool_name
unstructured_data_mcp_tool_errors_total       — Counter  — attributes: tool_name, error_type

Using otel.Meter("mcp.tools") — no-op if telemetry is disabled.

Instrument each tool handler

Files: list_pipelines.go, get_chunks.go, get_processed_document.go

Pattern — named return values + defer (Go equivalent of Python's @instrument_mcp_tool):

func(ctx context.Context, ...) (result *mcp.CallToolResult, aux any, retErr error) {
    start := time.Now()
    defer func() {
        status := "success"
        if result != nil && result.IsError { status = "error" }
        toolInvocations.Add(ctx, 1, metric.WithAttributes(
            attribute.String("tool_name", "..."),
            attribute.String("status", status),
        ))
        toolDuration.Record(ctx, time.Since(start).Seconds(), metric.WithAttributes(
            attribute.String("tool_name", "..."),
        ))
    }()
    // ... existing handler code unchanged ...
}

Existing handler logic is not modified — only the function signature changes to named returns and a defer block is added.

Verification

• make lint passes
• make test passes
• With ENABLE_OTEL=true OTEL_METRICS_EXPORTER=console, call a tool → see unstructured_data_mcp_tool_invocations_total and unstructured_data_mcp_tool_duration_seconds in terminal output

Depends on

• [Observability 1/10] Add OTEL metrics foundation — telemetry package + main.go wiring #308 (telemetry foundation)