Loki Mode

Verification depth + enterprise compliance surface (P1-2, P1-3, P3-6, P3-11)

• Held-out checklist for small specs (P1-2) (autonomy/prd-checklist.sh):
  small checklists (2 <= N < 4) previously reserved ZERO held-out items, leaving
  the entire checklist gameable. They now reserve exactly 1 held-out item via the
  same deterministic sha256 ranking used for larger sets. N=1 stays a no-op
  (cannot hold out from a single-item checklist); N>=4 is unchanged. Idempotent.
• Semantic test-authenticity gate wired on both routes (P1-3): the semantic
  detector (tests/detect-semantic-test-problems.sh, which catches fake tests
  that regex detectors miss) is now an opt-in quality gate via
  LOKI_GATE_SEMANTIC_TESTS (default off). Wired on the bash route
  (autonomy/run.sh) and mirrored on the Bun route
  (loki-ts/src/runner/quality_gates.ts) with byte-identical blocking semantics:
  blocks only on CRITICAL/HIGH findings; clean/absent/timeout/malformed never
  fires (deny-filtered); default-off means the detector is not invoked (zero
  cost, cannot deadlock the loop). Documented as the opt-in 9th gate in
  skills/quality-gates.md.
• Sandbox default kept opt-in, honestly documented (P3-6): sandbox mode is
  Docker-backed and hard-errors without Docker, so defaulting it on would break
  non-Docker users. It stays opt-in (LOKI_SANDBOX_MODE=true); the
  loki start CLI help now accurately states "default: off; requires Docker"
  instead of an ambiguous description.
• Continuous compliance surface (P3-11): new GET /api/compliance dashboard
  endpoint returns the live SOC2/ISO27001/GDPR compliance report generated from
  the real audit chain (via src/audit/index.js getReport + a CLI shim),
  including the real verifyChain() tamper-evidence verdict. Auth/tenant-scoped
  (require_scope("audit")); returns an honest empty report (never a fabricated
  "compliant" verdict) when no audit data exists, and degrades to
  available:false if the audit engine is unavailable.

Gates: local-ci 78/78, full pytest + bun test (78/78) + bash/Bun parity (11
shared toggles, 0 fail) green, held-out evals 34/34, 3-reviewer council
unanimous APPROVE.