v1.1.6 - Enterprise upload content-safety gate

Hardened hep-upload against malicious agent packages.

Detection (obfuscation-resistant, multilingual):

• Normalized detection shadow — NFKC, invisible/format-char stripping, Unicode Tag-block inline decode, homoglyph + leetspeak folding, separator collapse
• Boundary-free squished shadow for separated-letter tricks (i.g.n.o.r.e)
• Multi-line sliding window for injections split across lines
• Concepts in English, Korean, Chinese/Japanese, and major European languages
• Secret-exfiltration beacons and high-value credential access

Quality preservation (never blindly delete):

• Two-tier gate — redact only high-confidence attacker directives; keep + flag ambiguous, negated, quoted, or descriptive matches (security-training, prompt-engineering, devops-docs agents that legitimately use these terms)
• Packages still publish as ready with offending lines removed

Verified: 139 adversarial vectors across ~25 evasion families → 100% stripped; 35 realistic benign samples → 0 false positives. Full suite 415 passing.