fix(agentchat): respect allow_repeated_speaker=False in selector fallback

[extrasmall0]

Fixes #7471

Problem

When allow_repeated_speaker=False and the model exhausts max_selector_attempts without picking a valid (non-previous) speaker, the fallback in _select_speaker returns self._previous_speaker:

if self._previous_speaker is not None:
    trace_logger.warning(...)
    return self._previous_speaker  # ← returns the excluded speaker

This creates a livelock: agent A speaks → selector tries to pick someone else → exhausts attempts → falls back to A → A speaks again → repeat.

Root Cause

The _select_speaker method receives a participants list that already has the previous speaker filtered out (done by the caller when allow_repeated_speaker=False). The fallback logic ignores this filtered list and reads directly from self._previous_speaker.

Fix

Add an early-exit branch before the existing fallback: when allow_repeated_speaker=False and there is a previous speaker, use random.choice(participants) (the already-filtered list) instead of returning the excluded speaker.

if self._previous_speaker is not None and not self._allow_repeated_speaker:
    # participants already excludes the previous speaker
    return random.choice(participants)

The existing fallback for allow_repeated_speaker=True (return previous speaker) is unchanged.

[extrasmall0]

Quick follow-up: this is ready for review, and I wanted to add one concrete bit of context in case it helps triage. The regression happens in the selector fallback path when allow_repeated_speaker=False: after candidate filtering, the fallback can still pick a previously seen speaker instead of respecting the no-repeat setting. This patch keeps the existing selector flow intact and only fixes that fallback behavior. Happy to add a narrower regression test or trim the change further if that would make review easier.

[extrasmall0]

Quick follow-up after another week in queue: this is still ready for review, and the scope remains intentionally narrow. The change only fixes the selector fallback path when allow_repeated_speaker=False, so repeated speakers cannot slip back in after candidate filtering. There are no functional changes outside that fallback path. If it helps triage, I can split out the regression test or trim the patch further.

[extrasmall0]

Quick follow-up after rechecking this against current main: the patch is still a small, isolated fallback fix in the selector path. It only changes the exhausted selector fallback when allow_repeated_speaker=False, so instead of returning the previously excluded speaker it chooses from the already-filtered participant list. If it helps review, I can also add a focused regression test around the exhausted-selector path.