fix: skill metadata truncation must fall back to UTF-8 sequence boundary

[LIUXIN557]

Behavior change

skill_loader's truncate() used raw-byte substr(0, n-3) + "...". When a skill's description / whenToUse exceeds the 1024-byte budget and the cut lands inside a multi-byte UTF-8 character (long Chinese descriptions, e.g. lark-apps at 1289 bytes), the result is an invalid sequence (orphan lead byte + 0x2E from "...").

That string enters the skills index message; nlohmann::json body.dump() then aborts the entire request with:

[Error] [json.exception.type_error.316] invalid UTF-8 byte at index 16944: 0x2E

Trigger conditions (all reproduced):

• TUI only (headless does not inject the skills index message)
• Large-context model presets (e.g. deepseek-v4-flash with 1M window -> full descriptions included; 128K presets truncate before the bad bytes)
• Any skill with a >1024-byte description whose byte 1021 sits mid-character

Fix

• Route truncate() through the existing truncate_utf8_prefix() helper so cuts fall back to a UTF-8 sequence boundary.
• Guard final metadata with ensure_utf8() for defense in depth.

Tests

• New tests/skills/skill_loader_utf8_test.cpp: description and whenToUse truncation with a multi-byte char at the budget boundary must stay valid UTF-8.
• End-to-end: pty-driven TUI + --resume on a deepseek session reproduced type_error.316 deterministically before the fix; after the fix the request serializes and sends normally.

Verification

cmake --build build/macos-x64-release --target acecode acecode_unit_tests
ctest / gtest filter SkillLoaderTruncation.*  # 2/2 passed