fix: keep the leading system prompt when sanitizing an unanswered history

[BrennanTM]

Summary

Fixes #372. Since #339, _sanitize_history drops the leading system prompt from any history that has no reply yet, so the first request of every ChatAgent conversation started through format_query (the TUI path for its default agent) and any caller-supplied leading SystemMessage reach the model without it. The "unanswered trailing prompts" step computes the tail from the last AIMessage or ToolMessage; with no reply the tail is the whole list and only the last HumanMessage survives. Present in v0.17.1 and v0.17.2; absent in v0.17.0.

The change

A leading SystemMessage is held out of the tail computation and put back in front: head + body[:last_reply + 1] + kept prompt. Everything #339 intended is preserved: duplicate trailing prompts still collapse to the last one, now also on a first turn when a caller holding its own state retries, and mid-list system messages are still dropped. A history holding only a system prompt is returned as is instead of becoming an empty request. The hunk in base.py is six insertions and three deletions, with no signature change and no change to any caller.

Tests

tests/agents/test_sanitize_history.py, committed red first and flipped by the fix with no test edits. Four fail on main: the first ChatAgent request keeps its system prompt (it was human only), a caller-supplied leading SystemMessage survives the first call, a first-turn duplicate prompt collapses while the prompt stays, and a system-only history is left alone (it became empty). Three guards pass throughout: trailing duplicates after a reply still collapse, mid-list system messages are still dropped, and a history without a system prompt keeps only its last prompt. Running the test commit without the fix commit gives exactly 4 failed / 3 passed. Full suite locally with the dev dependencies installed: 1124 passed, 6 skipped, 1 xfailed (the pinned #295 case).

[BrennanTM]

CI note: the Windows job's single failure is tests/cli/tui/event_cards/test_plan.py::test_plan_card_renders_drafting_and_collapsed_steps (assert 'truncated' not in ... after the terminal resize), a TUI plan-card rendering test this PR does not touch. The same test failed with the same assertion on windows-latest in both CI runs of the docstring-only #368 (runs 38648069797 and 38791751260), while ubuntu and macos passed each time, and the latest run on main is green on all three. The rest of the Windows run here is 1124 passed, 5 skipped, 1 xfailed, including the seven new tests. I do not have permission to re-run the job. Happy to look at the flake separately if that would help.