hermes - 💡(How to fix) Fix guardrail-halt-style silent stream close at partial_stream_recovery + fallback_prior_turn_content sites [4 pull requests]

Context

Follow-up from #30770 / PR #31448 (which fixed the guardrail_halt case).

The fix added stream_delta_callback(final_response) to the guardrail-halt branch of agent/conversation_loop.py so SSE/TUI clients receive the synthesized halt message before the stream closes. Two structurally-identical sibling sites have the same gap and weren't touched by the original fix:

1. partial_stream_recovery (~L3556) — final_response = _recovered (text reconstructed from a partial stream after a think block). The recovered text wasn't necessarily streamed in full this turn.
2. fallback_prior_turn_content (~L3582) — final_response = agent._strip_think_blocks(fallback).strip() where fallback is the previous turn's content. The previous turn's content was streamed on the previous SSE response, so the current SSE writer drains an empty queue.

Both sites: assign final_response to text that wasn't (fully) streamed this turn, then break out of the loop. SSE clients see a finish chunk with zero content delta.

Impact

Lower than the guardrail case — these paths trigger on partial-stream / empty-response recovery rather than on every guardrail hit. But the failure shape is identical: blank Open WebUI bubble, indistinguishable from a crash.

Fix sketch

Mirror the guardrail-halt pattern: after assigning final_response and before break, fire it through agent.stream_delta_callback(final_response) if the callback exists. Optionally extract into a small helper since this is now a recurring pattern.

Why a separate issue

Keeping #31448 narrowly scoped to the reported bug. The recovery-path versions are less frequently hit and worth a small extraction pass (helper to dedupe the 3 emit sites) rather than a third inline copy of the same block.