fix(compressor): prevent orphan user turn after compaction (turn-pair preservation)

[H2KFORGIVEN]

Problem

When context compression fires, _ensure_last_user_message_in_tail anchors
cut_idx to the last user message so the active task isn't lost. However
_find_tail_cut_by_tokens then applies max(cut_idx, head_end + 1) before
returning — this advances the cut past the user message when
last_user_idx <= head_end, placing the user turn in the compressed region
without its assistant reply.

The LLM summariser sees an unanswered request and writes it into
## Pending User Asks. The next session's model then finds a dangling user
turn after the summary, treats it as a new instruction, and re-executes the
already-completed task.

Real observed symptoms:

• Agent turned lights off twice (once in session, once after compaction in a new session)
• Agent re-sent a message that was already sent
• Agent repeated a file-write that was already done

This is a latent second form of the original #10896 bug: the first fix kept
the user message out of the summary, but the max() call could silently undo
that guarantee.

Fix

Apply the Causal Coupling principle from agent-scaffolding literature:

A compaction boundary must never split a (user → assistant → tool_results)
turn-pair. Both halves must land on the same side of the cut.

Two additions:

1. _find_turn_pair_end(messages, user_idx)

Returns the index after the complete turn-pair that starts at user_idx:
user → assistant [→ tool results]. Stops at the next user message so it
never crosses into a different turn.

2. Causal Coupling guard in _ensure_last_user_message_in_tail

After computing adjusted = max(last_user_idx, head_end + 1), checks whether
the cut would land between the user and its reply. When a split is detected,
pushes cut_idx to pair_end so the whole pair lands in the compressed
region and is correctly summarised as completed.

Tests

All 81 existing compressor tests pass unchanged.

New class TestTurnPairPreservation (8 tests):

• 4 unit tests for _find_turn_pair_end (user-only, with reply, with tools, multi-turn)
• 4 tests for _ensure_last_user_message_in_tail (already-in-tail, pullback, orphan-prevention, end-to-end no-orphan invariant)

89 passed in 4.08s

Impact

• No behaviour change when the user message is already safely in the tail
• No performance impact — _find_turn_pair_end is O(k) where k is the number of tool results in the turn (typically 1–5)
• Fixes silent repeated-work bugs that are difficult to reproduce deterministically

[teknium1]

Merged via PR #56121: #56121

Your fix was reimplemented on current main (the branch was ~6000 commits stale and the code had moved from line 1190 to 2332), with your authorship preserved in git history on commit fc2fac7. The Causal Coupling guard and _find_turn_pair_end landed as you designed them, plus your 8 tests. Verified E2E: the orphan reproduces exactly as you described, and the fix keeps the completed turn-pair together. Thanks for the well-analyzed fix!