Yes, I was wrong earlier, I did not study the subject properly.
So, let me make a couple of assumptions.
1. $w$ is a string of finite length, so $|w| < \infty$
2. $w$ is a string of a language over finite alphabets, so $w \in \Sigma^{*}_{(a,b)}$
Now,
let me define the problem as
$L = \{<M, w> | \text{M is a TM and M halts on turning it's head left at least once during it's computation on } w, w \in \Sigma^*\}$
So, we can construct a decider as
1. Simulate $w$ on $M$
2. If $M$ turns it's head left, accept and accept.
3. Now, it might happens that the machine never turns left, in this case, we have following options -
(a) on reading current input alphabet $a,b$, machine goes right.
(b) machine has finished reading the input $w$, so it reads blank symbol $\$ \not \in \Sigma^*$, from the tape and moves right.
4. So, even if the machine is non-halting, it's state-transition diagram has finite transitions.
it repeats one of the above (a) or (b), which could be detected in at most $2 \cdot |w|$ steps which means we have a cycle in the state transition diagram.
This is where my assumptions come into picture.
If the input symbols were not finite, then I could have mapped them to natural numbers and then I could have had a machine that on reading a symbol, writes a random alphabet and goes right. Then I won't be able to tell if the machine is repeating the state or not.
If the length of the string was infinite, then well, I'll never be able to reject the string.
5. So, I can reject the string after finite number of steps.
Hence the $L$ is decidable.