fix: remediate 13 reachable security findings (cmdi, SSRF, LLM PII, DLP, secret, error-leak, CVE)

[Copilot]

Addresses a batch of 13 production-reachable security findings across the handler layer spanning command injection, SSRF, LLM prompt injection, PII leakage, a committed secret, internal error exposure, and a vulnerable dependency.

Changes

cwe.go — OS command injection

• DiagnosticPing: replaced exec.Command("sh", "-c", "ping -c 1 "+host) with AllowedHostname validation + exec.Command("ping", "-c", "1", host) (fixed argv, no shell)
• Errors logged server-side; generic message returned to client

ai.go — LLM prompt injection + error leak

• AIAnswer / AIAgentPlan: user-controlled input moved to user role; system instructions stay in system role — eliminates direct concatenation of untrusted data into the system prompt
• Removed internal prompts/tool specs from response body; err.Error() replaced with generic "invalid request"

dlp.go — PII in logs and HTTP POST

• SupportExport: SSN/DOB redacted in log.Printf (***-**-XXXX / YYYY-**-**); raw PII fields removed from the outbound analytics POST payload

secrets.go — committed GitHub PAT

• Removed hardcoded ghp_* constant; CloudTokens now reads REACH_TESTBED_GITHUB_TOKEN from the environment, returning HTTP 503 when unset

suspicious.go — SSRF + cleartext + error leak

• FetchTool: rejects non-https schemes; validates hostname against AllowedHostname allowlist before making any outbound request; all err.Error() calls replaced with generic messages + server-side log.Printf

cve.go — error leak

• ParseLanguage: returns "invalid language tag" instead of err.Error()

go.mod / go.sum — CVE dependency

• Upgraded golang.org/x/text v0.3.7 → v0.3.8 to resolve grouped CVEs