Mid-turn compaction no longer drops this turn's tool results from the next request (salvage #91234)

[teknium1]

Mid-turn compaction no longer drops the current turn's earlier tool calls and results from the next request, so what the model receives matches state.db again.

• Post-tool compaction (salvaged from fix(agent): re-anchor after post-tool compression #91234 by @Diaspar4u, authorship kept): compress_after_tool_results re-anchors current_turn_user_idx with _reanchor right after a real compaction and hands it back through the post-tool and tool-round verdicts. _run_phase copies it into the loop state, and _persist_user_message_idx updates at the same moment.
• Every other mid-turn site (ours): prepare_iteration checks the index before every request. It must land on this turn's user row, either verbatim or through its user-originated view. If it doesn't, it re-anchors. This covers the pre-API pressure gate (run_preflight_compression returns continue with the pre-compaction index) and any compaction site added later. Where fix(agent): re-anchor after post-tool compression #91234 already re-anchored, the check does nothing.
• Tests: tests/agent/test_mid_turn_compaction_turn_boundary.py holds 2 invariants: fix(agent): re-anchor after post-tool compression #91234's post-tool test (moved into this file) and a real AIAgent loop test where the pre-API gate compacts after a tool round.

Root cause: build_api_messages splits the replay prefix at current_turn_user_idx (canonicalize_replay_history(messages[:idx])). Mid-turn compaction shrinks messages from the front, so a stale index lands inside this turn's tool rows. Canonicalization drops the assistant tool_call whose result fell past the split, and the sanitizer then drops the orphaned result.

Live repro (real AIAgent + on-disk SessionDB + real read_file tools against the loopback fake provider; tests/e2e/core/compaction/):

• Before (origin/main 07646a7f727): test_compaction_auto.py::test_good_summarizer...[11] fails with turn 1 request 2: request history diverges from state.db. sent=[… U('U11x1…'), A[call_fake_5], T:call_fake_5] vs db=[… U('U11x1…'), A[call_fake_4], T:call_fake_4, A[call_fake_5], T:call_fake_5]: the call_fake_4 pair is missing from the wire.
• After: all 34 tests in tests/e2e/core/compaction/ pass (4 files, including the kill -9 and summary-fault matrices).

invariant test	origin/main	#91234 only	this PR
post-tool compaction re-anchors the boundary	red (TypeError, no kwarg)	green	green
pre-API compaction mid-turn keeps this turn's tool pair on the wire	red (request ends at the user row, call_1 pair dropped)	red	green

That middle column is why the prepare_iteration check is layered on top of the salvage.

Validation

• scripts/run_tests.sh tests/agent/: 9219 passed, 1 failed, then fixed. The failure was test_proactive_prune_loop_wiring.py calling compress_after_tool_results without the new required current_turn_user_idx kwarg; the call now passes it and no assertion changed. test_run_agent.py hit the 300 s per-file cap under host load and passed on rerun (267 tests), along with test_moa_loop_mode.py, which failed once and passed on retry and is unrelated to this change. After rebasing onto 33a30fdd815: the touched test files are green.
• ruff check on the touched files: clean. scripts/check_no_tmp_literals.py agent tests/agent: clean. git diff --check: clean. scripts/audit_pr_attribution.py --fix: all mapped.
• The E2E suite that found this lands separately in the core E2E-suite PR and is not part of this diff.

Credit / Supersedes

• Supersedes fix(agent): re-anchor after post-tool compression #91234 (@Diaspar4u): its commit is cherry-picked here, and this PR adds the pre-API gate it did not cover.
• Credit fix: update current_turn_user_idx after preflight compression #39239 (@moizogg), the first report of this stale index (preflight compression). main already re-anchors turn-start preflight, and the prepare_iteration check covers the per-iteration pre-API gate its thread asked about.
• Credit fix: recalculate memory prefetch injection index after message sequence repair #43552 (@cwchewang), the same stale index after alternation repair, which main has since fixed (fix(loop): the current-turn index is re-anchored after the alternation repair merges rows (webui #7461; salvage #105694) #106312). Its "re-derive the anchor right before request build" suggestion is the shape used here.
• Related, not superseded: fix(agent): keep pre-API /steer off prior-turn tool results #107274 (steer placement uses the same index as a floor; a correct index makes it safer) and feat(loop): turn-boundary contract v2 — mark this turn's user row and export it on every result envelope #108216 (turn-boundary contract export).

Infographic