Fix inverted use_stale_model gating in DDWEStreamThinker

[braceal]

The two branches guarding inference submission in DDWEStreamThinker.process_simulation_result were swapped relative to what use_stale_model means:

if not self.use_stale_model:
    while self.train_output is None:      # only waits for *a* model
        time.sleep(10)
elif self.use_stale_model:
    while self.train_iteration < self.ensemble.iteration:   # waits for a *fresh* model
        time.sleep(10)

So use_stale_model=False — the strict setting — merely waited for train_output to become non-None and would then submit inference against an arbitrarily old model. use_stale_model=True — the permissive setting — took the strict path and blocked until train_iteration caught up with the current ensemble iteration, serializing inference behind training on every iteration.

The branches are now swapped so the flag means what it says:

• use_stale_model=True blocks only until the first model exists, then reuses whatever is current.
• use_stale_model=False blocks until the current iteration's model is ready.

Log messages are updated to distinguish the two waits ("first" vs "next" streaming train task), which previously printed identical text from both branches.

Extracted from #40 so it can land on its own.

🤖 Generated with Claude Code