fix(agent): clamp tail-cut boundary and summary-scan indices to prevent IndexError

[x7peeps]

背景

修复 #75588: 当短对话以 tool-call/result 组结尾且受保护的 head 对齐到消息列表末尾时，_find_tail_cut_by_tokens() 会返回 len(messages)+1，导致 _find_context_summaries() 在索引 messages[idx] 时触发 IndexError，直接中断 gateway 活跃轮次。

根因分析

当 head_end >= len(messages) 时，_find_tail_cut_by_tokens() 末尾的 return 语句使用 max(cut_idx, head_end + 1) 会将索引推到 len(messages)+1。这个越界值随后传入 _find_context_summaries()，后者在 range(start, end) 中迭代并直接索引 messages[idx]，未做任何范围钳制，最终抛出 IndexError。

修复方式

两层防御：

1. 源头修复：_find_tail_cut_by_tokens() 的返回值用 min(n, ...) 钳制，确保永远不超过 len(messages)
2. 防御层：_find_context_summaries() 对 start/end 参数做 [0, len(messages)] 范围钳制，即使未来调用方传入越界值也不会崩溃

验证

• 新增 7 个回归测试，覆盖精确的越界场景
• 全部 214 个现有 test_context_compressor.py 测试通过
• 遵循项目 AGENTS.md 贡献规范
• 无破坏性变更

Closes #75588

[teknium1]

Thanks for the focused boundary fix. Current main still has the invalid tail-cut return: agent/context_compressor.py:4940 can return len(messages) + 1 when head_end == len(messages), and _align_boundary_forward() preserves that out-of-range value (agent/context_compressor.py:4408-4410). The proposed min(n, ...) clamp makes the existing no-compressible-window guard in compress() (agent/context_compressor.py:5095-5119) handle that state correctly.

Suggested changes

• Add an end-to-end compress() regression for the aligned-head-at-end shape. Patch _generate_summary and assert it is not called and the transcript is unchanged. The new tests directly cover the helpers, but the production boundary path is agent/context_compressor.py:5074-5119.

Automated hermes-sweeper review.

[teknium1]

Please add a compress()-level regression for this aligned-head-at-end transcript, asserting _generate_summary is not called and the input is returned unchanged. This direct helper assertion does not exercise the production compress_start >= compress_end no-op branch.

[teknium1]

Merged via salvage PR #75884 (#75884) — your commits were cherry-picked onto current main with authorship preserved via rebase-merge, plus a compress()-level E2E regression test on top. Fixes #75588. Thanks for the precise two-site diagnosis and tests.