fix(compression): keep lite from truncating the latest tool results

[QuangBlue]

Problem

The default compression combo ("Standard Savings": session-dedup → lite) runs for every key with compression enabled. lite.compressToolResults cut every role:"tool" string to 2000 chars (…[truncated]), including the results of the tool calls the model had just made.

The model never saw past the cut, so it read the file again. The re-read was cut again, or deduped by session-dedup into a [dedup:ref] pointing at the cut copy, and OpenAI-format agents (Cursor IDE, Kilo/OpenCode) looped on the same read. A Cursor IDE session on our deployment re-read the same two files 15+ times; every result was 2007–2015 chars ending in …[truncated], including the newest.

Anthropic-format tool_result blocks are not affected, since compressToolResults only handles string role:"tool" messages.

Fix

Skip tool messages after the last assistant message, which are the results of the latest tool calls. This is the same current-turn rule session-dedup already applies. Older turns are still truncated. A body with no assistant message keeps the previous behaviour.

Tests

tests/unit/compression/lite-current-turn.test.ts:

• the newest tool result stays whole;
• earlier-turn results are still truncated;
• a re-read stays whole through the stacked session-dedup + lite pipeline.

Reproduced and verified against a real model through a local router with an OpenAI-format agent that uses a Kilo-style read tool, and with Cursor IDE's own system prompt and tool set replayed. Before the fix, 3/3 runs looped 12 reads without an answer, and 0/3 answered with the Cursor IDE shape. After the fix, 3/3 answered correctly in 1–2 reads.

The compression suites show the same failures as the base tip (previewRouteFuzzy fuzzyDedup CCR marker, #7849 heap test). Neither is touched by this change.

⚠️ base-red inherited: #15032