feat(formula-safety): add opt-out to run unsafe formulas

[sommaa]

Summary

Adds a global, persisted (default-off) "Allow unsafe formulas" setting that disables the formula sandbox end-to-end, so trusted templates with non-whitelisted formulas can be imported, saved, and rendered. On a desktop-first, single-user, local app the user owns the machine and may knowingly trust their own templates.

Both enforcement layers honor a single runtime flag (otherwise a template would import but still fail to render):

• Eval-time — safe_eval / safe_exec / assert_formula_safe / assert_no_escape in formula_safety.py
• Template-time — assert_template_formulas_safe + the validate loop in templates.py

Changes

• formula_safety: runtime flag (is_unsafe_allowed / set_unsafe_allowed) seeded from config; gate functions no-op and safe_eval/safe_exec run with real builtins when enabled
• config: allow_unsafe_formulas seed (default false, env-overridable)
• templates API: skip template-time formula checks when enabled
• settings API (new): GET/PUT /api/v1/settings/security to toggle the flag at runtime
• frontend: Security toggle in SettingsMenu, persisted in localStorage and mirrored to the backend (re-pushed on startup); "Enable unsafe formulas & retry" prompt in TemplateManager on unsafe-import rejection (covers the onboarding wizard too)
• tests: backend coverage for the flag behavior and the settings endpoint

Security note

This is a deliberate opt-out: once enabled, a malicious template becomes arbitrary code execution. It defaults off, and both the Settings toggle and the import prompt spell out the risk.

Verification

• 13 new backend tests + full suite (1031 passed; only pre-existing parquet/pyarrow failures, unrelated)
• Frontend production typecheck: 0 new errors; uiSlice runtime tests pass
• Driven end-to-end through the real :3000 → :8000 UI stack: unsafe template blocked (400) while off, accepted (200) after enable & retry, blocked again after reset

🤖 Generated with Claude Code