fix(context-compact): fix L1 truncation bug on consecutive tool results

[lelewang520]

Summary

Fixes a hidden context truncation bug in the L1 snip_compact pipeline where consecutive tool_result messages (often generated by parallel tool execution) would cause a validation failure, leading to broken tool/result pairings and API crashes.

The Problem

In the original snip_compact implementation, the tail boundary alignment (tail_start) was guarded by a single if statement:

if (tail_start > 0 and tail_start < len(messages)
        and _is_tool_result_message(messages[tail_start])
        and _message_has_tool_use(messages[tail_start - 1])):
    tail_start -= 1

[Bill-Billion]

Thanks for finding the consecutive tool_result boundary case.

#470 covers this tail-boundary fix and also corrects max_messages accounting and small-limit behavior. The multi-result case handled here is included there, so we’ll keep the work in one PR and close this one as superseded.

Thanks for the focused fix and explanation.