fix(agent): repair transcript pairing after compaction and truncation

[daotri]

Problem

When a conversation is compacted mid-loop or truncated by the background
summarizer, the split point could land in the middle of a tool-chain,
leaving an orphaned role:"tool" message with no preceding
assistant message carrying tool_calls. Providers reject this — OpenAI
returns a 400 on the next request.

Fix

Repair transcript pairing at both truncation points, addressing the
root cause rather than a single call site:

1. Mid-loop compaction — the compacted slice (persisted verbatim as run
   history) now runs through sanitizeHistory, so orphaned tool messages
   never reach the provider. A slog.Warn("post_compaction_sanitize", ...)
   surfaces when a repair happens.
2. Background summarization — truncation now walks back to a clean
   tool-chain boundary via a shared toolChainSplitIndex helper, so the kept
   tail never starts mid tool-chain.

The backward-walk logic (previously inlined in compaction) is extracted into
toolChainSplitIndex and reused by both paths, with an added bounds guard
(splitIdx < len(messages)) that also fixes a latent out-of-range access when
the split index equals the message count.

Changes

File	Change
internal/agent/loop_compact.go	Extract shared toolChainSplitIndex helper
internal/agent/loop_history_sanitize.go	Summarizer truncates on tool-chain boundary
internal/agent/loop_pipeline_callbacks.go	Sanitize compacted history before persist
internal/agent/loop_compact_boundary_test.go	Table-driven unit tests (new)

Testing

• go build ./internal/agent/ ✅
• go vet ./internal/agent/ ✅
• go test ./internal/agent/ -run 'TestToolChainSplitIndex|TestMakeCompactMessages_SanitizesOrphanTool|TestMaybeSummarize_ToolChainBoundary' ✅

New unit tests cover: clean boundary stays put, split on tool result walks
back before assistant(tool_calls), consecutive tool results walk the whole
chain, compaction callback sanitizes orphans, and summarizer truncates on the
tool-chain boundary.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Eeg3HGEWSqhPxsSQKAbe2x