feat(security): response verification layer for post-LLM injection detection (#1862)

[bug-ops]

Summary

Implements the missing third layer of Zeph's defense-in-depth pipeline per arXiv 2511.15759 (SecureAgent benchmark). Reduces attack success by adding post-LLM response verification before tool dispatch.

• Layer 1 (existing): ContentSanitizer — scans untrusted input before context insertion
• Layer 2 (existing): ExfiltrationGuard + PreExecutionVerifier — output filtering + pre-dispatch tool audit
• Layer 3 (this PR): ResponseVerifier — scans LLM responses for echoed injection patterns

Changes

• zeph-sanitizer: new ResponseVerifier struct with verify(ctx) -> ResponseVerificationResult
• zeph-tools/patterns.rs: 7 curated RAW_RESPONSE_PATTERNS (separate from input-side patterns to avoid false positives)
• zeph-config: ResponseVerificationConfig with enabled=true, block_on_detection=false defaults
• zeph-core agent loop: called in native and legacy tool-execution paths after LLM response, before tool dispatch
• TUI SEC panel: [rver] badge via new SecurityEventCategory::ResponseVerification variant
• 10 unit tests covering all pattern categories and Clean/Flagged/Blocked result paths

Test plan

• cargo +nightly fmt --check — clean
• cargo clippy --workspace --features full -- -D warnings — clean
• cargo nextest run --workspace --features full --lib --bins — 6407 passed (+10 vs baseline 6397)
• All 10 response verifier tests pass
• raw_response_patterns_all_compile test guards against regex regressions

Closes #1862