Which runtime data areas are per-thread versus shared, and why does that distinction matter?

The heap and Metaspace are shared across all threads, while each thread gets its own JVM stack, PC register, and native stack.

That is precisely why heap data needs synchronization and the memory model, but local variables and the call stack are inherently thread-safe and never escape a thread unless you publish them. It also explains thread cost: each platform thread reserves stack memory (often ~1 MB by -Xss), so tens of thousands of threads exhaust native memory long before the heap does.