`src/compact/truncating.rs` — `TruncatingCompactor` and `TokenSplitter`

The default shrink strategy (dependency-free, no LLM calls) and the exported splitter for custom compactors.

Key items

• TruncatingCompactor { preserve_recent: 4, min_messages: 6 } + builders (clamped to minimums 1 / 2).
• TokenSplitter — same knobs; split(&messages) -> SplitResult { to_compact, preserved, token counts, split_index }.
• Tool-pair machinery (private): PartMate/ToolPairing — per-occurrence call↔result pairing, orphan-safe split adjustment, first-message result reattachment.

Behavior notes

• Rules: conversations under min_messages pass through; the first message is always kept; tool call/result pairs are never orphaned (the split point moves, a fixpoint loop rechecking newly admitted messages); never returns an empty list.
• The target token budget is a hint here — the truncator sheds by message count; the manager’s fit check decides adequacy.
• TokenSplitter splits only at assistant→user turn boundaries that are pair-safe; no safe boundary → everything is preserved (split_index: 0) — the safe direction.
• Even “no change” paths filter orphaned lone results and report the savings honestly.

Deep dive: Compaction.