feat(security): add Stage-2 LLM semantic scan for third-party skills

[bug-ops]

Summary

Implements defense against Semantic Compliance Hijacking (SCH) attacks (arXiv:2605.14460), where malicious third-party skills encode harmful behavior through natural language instructions in SKILL.md without any explicit code payload, bypassing static security scans.

• zeph-skills: new SkillSemanticScanner (semantic_scanner.rs) — LLM-based compliance check using chat_typed_erased with a configurable fast provider; 8 KiB content cap with head+tail sampling; XML delimiter-escape neutralization before prompt interpolation; ScanVerdict::{Allow, Warn, Block} with unknown-token fallback to Block
• zeph-plugins: scan_targets() method extracts SKILL.md candidates from archive before installation, keeping the crate LLM-free; includes canonicalize/starts_with path-traversal guard
• zeph-core: semantic_scan_plugin_add wires the scanner into handle_plugins; fail-closed on Block; returns config error when semantic_scan = true but semantic_scan_provider is empty
• zeph-config: doc comments corrected for semantic_scan, semantic_scan_provider, and SemanticViolation

Opt-in: [skill] semantic_scan = true + semantic_scan_provider = "<provider-name>" in config.

Test plan

• cargo nextest run --config-file .github/nextest.toml --workspace --lib --bins — 10341 tests pass (10339 existing + 2 new security tests)
• cargo clippy --workspace -- -D warnings — clean
• cargo +nightly fmt --check — clean
• RUSTFLAGS="-D warnings" cargo check --workspace --all-targets --features desktop,ide,server,chat,pdf,scheduler — clean
• LLM serialization gate: live API test required before merge (scanner uses chat_typed_erased); use local Ollama provider (OpenAI 429 gate currently blocked)
• Verify semantic_scan = true + valid provider → scan runs on zeph plugins add <url>
• Verify semantic_scan = true + empty provider → config error returned

Closes #3947