[DarknessKiller]

Problem

Lite's tool-result cap keeps only the head and throws the rest away, with no way for the model to know what the dropped region contained beyond a single marker. Any consumer that addresses content by line - an anchored edit tool, a line-numbered grep, a diff - then looks up a line that is no longer in the context.

Measured with the current defaults (maxToolLength = 2000):

tool result	in	out after lite
95-line numbered file read	5873 chars / 95 lines	2013 chars / 35 lines (cut after line 33)
rg output, 108 matches	9116 chars	2008 chars, 24 matches left

Everything the agent was about to edit is in the 60 dropped lines, and the model rebuilds them from memory instead.

Fix

truncateToolText now keeps the tail too for multi-line content: ~70% of the budget from the head, ~30% from the end, both on line boundaries, with an honest ... [N lines elided] ... marker in the middle. The existing ...[truncated] marker stays at the end, so every current consumer and test contract is unchanged.

Single-line content keeps the #8169 word-boundary cut (no tail slice, nothing to align).

Same fixture after the fix: 2057 chars - head through line 33, ... [61 lines elided] ..., tail through the last line.

Tests

• new tests/unit/compression/lite.test.ts case: 95 line-addressed lines, first and last line must survive, elision marker present. Fails before the fix with last line must survive the cut.
• tests/unit/compression/lite.test.ts + tests/unit/8169-lite-word-boundary-truncation.test.ts + tests/unit/compression/lite-current-turn.test.ts -> 43 pass.
• tests/unit/chatcore-compression-integration.test.ts + tests/integration/chatcore-compression-integration.test.ts -> pass.
• tests/unit/compression/*.test.ts (full dir) -> 1944 pass.

Scope

Only open-sse/services/compression/lite.ts and its test. No config, no env var, no default cap change.