feat(agent): add jittered retry backoff

Adds agent/retry_utils.py with jittered_backoff() — exponential backoff
with additive jitter to prevent thundering-herd retry spikes when
multiple gateway sessions hit the same rate-limited provider.

Replaces fixed exponential backoff at 4 call sites:
- run_agent.py: None-choices retry path (5s base, 120s cap)
- run_agent.py: API error retry path (2s base, 60s cap)
- trajectory_compressor.py: sync + async summarization retries

Thread-safe jitter counter with overflow guards ensures unique seeds
across concurrent retries.

Trimmed from original PR to keep only wired-in functionality.

Co-authored-by: martinp09 <martinp09@users.noreply.github.com>

trajectory_compressor.py

@@ -44,6 +44,7 @@ import fire
 from rich.progress import Progress, SpinnerColumn, TextColumn, BarColumn, TaskProgressColumn, TimeElapsedColumn, TimeRemainingColumn
 from rich.console import Console
 from hermes_constants import OPENROUTER_BASE_URL
+from agent.retry_utils import jittered_backoff
 
 # Load environment variables
 from dotenv import load_dotenv
@@ -585,7 +586,7 @@ Write only the summary, starting with "[CONTEXT SUMMARY]:" prefix."""
                 self.logger.warning(f"Summarization attempt {attempt + 1} failed: {e}")
                 
                 if attempt < self.config.max_retries - 1:
-                    time.sleep(self.config.retry_delay * (attempt + 1))
+                    time.sleep(jittered_backoff(attempt + 1, base_delay=self.config.retry_delay, max_delay=30.0))
                 else:
                     # Fallback: create a basic summary
                     return "[CONTEXT SUMMARY]: [Summary generation failed - previous turns contained tool calls and responses that have been compressed to save context space.]"
@@ -647,7 +648,7 @@ Write only the summary, starting with "[CONTEXT SUMMARY]:" prefix."""
                 self.logger.warning(f"Summarization attempt {attempt + 1} failed: {e}")
                 
                 if attempt < self.config.max_retries - 1:
-                    await asyncio.sleep(self.config.retry_delay * (attempt + 1))
+                    await asyncio.sleep(jittered_backoff(attempt + 1, base_delay=self.config.retry_delay, max_delay=30.0))
                 else:
                     # Fallback: create a basic summary
                     return "[CONTEXT SUMMARY]: [Summary generation failed - previous turns contained tool calls and responses that have been compressed to save context space.]"