fix(tool): keep truncated results within maxBytes for small caps

[Alan-Yu-2077]

Summary

applyResultSizeLimit() computed a head/tail budget with Math.max(0, maxBytes - suffixBytes) but still appended the full truncation suffix afterwards. When maxBytes was smaller than the suffix ("\n[Tool output truncated: head and tail shown.]", 46 bytes), the returned text exceeded the requested cap.

Per the report, a one-byte result with maxBytes = 0 returned the full 46-byte notice and metadata.returnedBytes recorded 46 — both larger than the requested maximum.

This fix caps the notice itself when maxBytes cannot fit the suffix, and otherwise reserves room for the suffix as before. The returned text and metadata.returnedBytes now stay within maxBytes for every non-negative limit. UTF-8 safety is preserved via the existing truncateUtf8 helper.

Fixes #426

Test plan

Extended tests/tool/tool-result-size-limit.spec.ts:

• returned bytes and metadata.returnedBytes never exceed maxBytes for 0, 1, suffixBytes - 1, suffixBytes, and suffixBytes + 1;
• a multibyte (€) result stays within the cap and contains no U+FFFD replacement characters at the byte boundary;
• the existing head/tail truncation test is unchanged.

$ node --test dist/tests/tool/tool-result-size-limit.spec.js
# tests 3
# pass 3
# fail 0

Full suite (npm test) stays green.