fix(agent): forceCompression recovers from single oversized Turn

When the entire session history is a single Turn (e.g. one user message
followed by a massive tool response), findSafeBoundary returns 0 and
forceCompression previously did nothing — leaving the agent stuck in
a context-exceeded retry loop.

Now falls back to keeping only the most recent user message when no
safe Turn boundary exists. This breaks Turn atomicity as a last resort
but guarantees the agent can recover.

Also updates docs/agent-refactor/context.md to document this behavior.

Ref #1490

docs/agent-refactor/context.md

@@ -103,7 +103,9 @@ This prevents wasted (and billed) LLM calls that would otherwise fail with a con
 
 `forceCompression` runs when the LLM returns a context-window error despite the proactive check.
 
-Drops the oldest ~50% of Turns. Stores a compression note in the session summary (not in history messages) so `BuildMessages` can include it in the next system prompt.
+Drops the oldest ~50% of Turns. If the history is a single Turn with no safe split point (e.g. one user message followed by a massive tool response), falls back to keeping only the most recent user message — breaking Turn atomicity as a last resort to avoid a context-exceeded loop.
+
+Stores a compression note in the session summary (not in history messages) so `BuildMessages` can include it in the next system prompt.
 
 This is the fallback for when the token estimate undershoots reality.