Empty LLM Response Retry at Pipeline Level

Problem Statement

When a provider returns empty content (resp.Content == "") with no tool calls and finish_reason: "stop", the pipeline treats it as a valid
final answer. The finalize stage substitutes "..." as a placeholder — user sees literal dots with no explanation. This is a silent failure
that degrades trust.

Root Cause Chain

Provider returns empty/malformed response
→ ThinkStage: resp.ToolCalls == 0 → BreakLoop (treats as final answer)
→ ObserveStage: FinalContent = "" (empty)
→ FinalizeStage: FinalContent == "" → "..." fallback
→ User sees "..."

Proposed Approach: Pipeline-Level Silent Retry

Core idea: Detect "empty response" as a distinct failure class in ThinkStage (like truncation is today), retry up to 2 times with a system
hint, and only fall through to a user-friendly error message if all retries fail.

Detection Criteria (all must be true)

An LLM response is classified as "empty" when:

• resp.Usage == nil or resp.Usage.TotalTokens == 0
• len(resp.ToolCalls) == 0 (no actions)
• resp.Thinking == "" (no reasoning output)
• NOT a deliberate NO_REPLY (that's a valid empty response for group chats — but NO_REPLY is detected by content matching, so empty string
  won't match)

Start by tracing the ThinkStage, ObserveStage, and FinalizeStage flow described in the issue, then inspect how truncation retries and NO_REPLY detection are handled. Add empty-response detection and bounded retries with a system hint, and verify that repeated empty responses produce a user-friendly error instead of the "..." fallback.