fix(core): sanitize orphaned tool_use/tool_result on history restore (#1360)

[bug-ops]

Summary

• Cross-session history restore could produce invalid tool_use/tool_result sequences at history boundaries, causing Claude API 400 errors on session resume
• Add sanitize_tool_pairs() post-load sanitization in load_history() that removes orphaned tool messages at both ends of restored history
• 6 unit tests added covering all boundary conditions

Root causes

RC-1: load_history_filtered() LIMIT clause can split a tool_use/tool_result pair at the boundary, leaving an orphaned tool_use as the last restored message.

RC-2: Session interruption (Ctrl+C, timeout, crash) between persisting the assistant tool_use message and the user tool_result message leaves an orphaned tool_use in SQLite. On next session restore this triggers a Claude API 400.

Changes

• crates/zeph-core/src/agent/persistence.rs: Add private sanitize_tool_pairs(messages: &mut Vec<Message>) -> usize. Called in load_history() after the loading loop on the just-loaded slice (split off from self.messages to exclude the system prompt). The function loops until stable, removing:
  1. Trailing assistant messages that have ToolUse parts with no following user ToolResult
  2. Leading user messages that have ToolResult parts with no preceding assistant ToolUse
     Each removal logs tracing::warn! with affected tool IDs.

Tests

6 new unit tests in agent::persistence::tests:

Test	Scenario
load_history_removes_trailing_orphan_tool_use	Trailing orphan removed
load_history_removes_leading_orphan_tool_result	Leading orphan removed
load_history_preserves_complete_tool_pairs	Valid pair preserved
load_history_handles_multiple_trailing_orphans	Multiple consecutive orphans removed
load_history_no_tool_messages_unchanged	Plain messages pass through
load_history_removes_both_leading_and_trailing_orphans	Loop handles both ends in one call

Test plan

• cargo +nightly fmt --check passes
• cargo clippy --workspace --features full -- -D warnings passes
• cargo nextest run --workspace --features full --lib --bins passes (4693 tests)
• All 10 load_history tests pass
• config_default_snapshot failure is pre-existing on main (confirmed), unrelated to this PR

Notes

• No SQL queries modified
• O(n) remove(0) for leading orphan removal is acceptable given history_limit bounds; can be optimized with VecDeque in a future pass (R-1)
• Mid-sequence orphan detection (RC-4) deferred as separate low-severity issue (R-2)

Closes #1360