Summary

When deferred tool pair summarization fires (batch-applied deferred tool summaries applied=N), it replaces tool_result messages with plain-text summaries but leaves the corresponding tool_use messages in the database. On the next session restore, these tool_use messages have no matching tool_result and are stripped as orphaned mid-history pairs.

Reproduction

1. Run a long session with tool_call_cutoff = 6 and many memory_save calls
2. Deferred summarization fires: tier-0: batch-applied deferred tool summaries applied=6 count_pressure=true
3. Start a new session — observe many WARN: stripping orphaned mid-history tool_use parts

Evidence

After the Alice session (CI-211, 19 turns, 9 memory saves):

applied deferred tool pair summaries count=6
tier-0: batch-applied deferred tool summaries applied=6 token_pressure=false count_pressure=true

Next session (CI-212):

WARN stripping orphaned mid-history tool_use parts tool_ids={"call_n8k3ZB450zOj9rhj6pqW1b9y"} index=2
WARN stripping orphaned mid-history tool_result parts tool_use_ids={"call_n8k3ZB450zOj9rhj6pqW1b9y"} index=4
... (12 pairs total stripped)

DB inspection confirms: tool_use message (id=83) contains both tool_use IDs; ids 84-85 are plain-text summaries (tool_result replaced); id=86 is a raw tool_result for only one of the two parallel calls.

Impact

• MEDIUM: Grows unboundedly — each session with deferred summarization adds more orphaned tool_use messages to the DB
• Every subsequent restore pays the cost of stripping N orphaned pairs
• Context history becomes increasingly mismatched between DB records and actual conversation

Fix Direction

When batch_apply_deferred_summaries() replaces tool_result messages with plain-text summaries, it must also:
a) Remove the corresponding tool_use messages from the DB (or mark them as summarized), OR
b) Update the tool_use messages to reference the summary instead of expecting a tool_result

Discovered CI-212 (2026-03-27)